CN113449200B - Article recommendation method and device and computer storage medium - Google Patents

Article recommendation method and device and computer storage medium Download PDF

Info

Publication number
CN113449200B
CN113449200B CN202010224857.7A CN202010224857A CN113449200B CN 113449200 B CN113449200 B CN 113449200B CN 202010224857 A CN202010224857 A CN 202010224857A CN 113449200 B CN113449200 B CN 113449200B
Authority
CN
China
Prior art keywords
entity
vector
items
triplet
determining
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010224857.7A
Other languages
Chinese (zh)
Other versions
CN113449200A (en
Inventor
吴小飞
浦世亮
姜伟浩
葛挺
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hangzhou Hikvision Digital Technology Co Ltd
Original Assignee
Hangzhou Hikvision Digital Technology Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hangzhou Hikvision Digital Technology Co Ltd filed Critical Hangzhou Hikvision Digital Technology Co Ltd
Priority to CN202010224857.7A priority Critical patent/CN113449200B/en
Publication of CN113449200A publication Critical patent/CN113449200A/en
Application granted granted Critical
Publication of CN113449200B publication Critical patent/CN113449200B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9536—Search customisation based on social or collaborative filtering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/25—Integrating or interfacing systems involving database management systems
    • G06F16/254—Extract, transform and load [ETL] procedures, e.g. ETL data flows in data warehouses
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • G06F16/367—Ontology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/06—Buying, selling or leasing transactions
    • G06Q30/0601—Electronic shopping [e-shopping]
    • G06Q30/0631—Item recommendations

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Business, Economics & Management (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Accounting & Taxation (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Finance (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Evolutionary Computation (AREA)
  • Strategic Management (AREA)
  • Marketing (AREA)
  • Animal Behavior & Ethology (AREA)
  • General Business, Economics & Management (AREA)
  • Economics (AREA)
  • Development Economics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The application discloses an article recommendation method, an article recommendation device and a storage medium, and belongs to the field of information recommendation. The method comprises the following steps: determining first prediction scores of a plurality of articles through a collaborative filtering model according to article behavior data of a target user; determining k first items from the plurality of items based on the first prediction scores for the plurality of items; determining first entity vectors of the k first items according to the k first items and the knowledge graph vector set; determining second prediction scores of the k first items according to the first entity vectors of the k first items and item behavior data of the target user; and if the mean square error between the first prediction scores and the second prediction scores of the k first items is less than or equal to the error threshold value, recommending the k first items to the target user. The recommendation algorithm fully considers the similarity between the article attributes on the basis of determining the article similarity according to the article behavior data, and further improves the recommendation accuracy.

Description

Article recommendation method and device and computer storage medium
Technical Field
The present application relates to the field of information recommendation, and in particular, to a method and an apparatus for recommending an article, and a computer storage medium.
Background
Currently, network platforms can provide users with online recommendation services for items such as news, goods, pictures, video, audio, documents, and the like. When a user browses, collects or scores articles on the network platform, the server of the network platform records article behavior data of the user, and mines user preferences according to the article behavior data of the user to recommend articles of interest to the user.
In the related art, a collaborative filtering recommendation algorithm is generally adopted to recommend an item for a user. Specifically, for a target user on the network platform, the article behavior data of the target user can be acquired, then the prediction scores of a plurality of articles in the article set are determined through the collaborative filtering model according to the article behavior data of the target user, and some articles with higher prediction scores are selected from the plurality of articles and recommended to the user according to the prediction scores of the plurality of articles. The forecasting grade of each item is used for indicating the interest degree of a target user for each item, and the collaborative filtering model is used for determining the similarity matrix of a plurality of items in the item set according to the item behavior data of a large number of users on the network platform, and then determining the forecasting grade of the plurality of items in the item set according to the item behavior data of the target user and the similarity matrix of the plurality of items. Wherein the similarity matrix of the plurality of items comprises the similarity between two items in the plurality of items.
However, the collaborative filtering recommendation algorithm analyzes the similarity between the articles according to the behavior data of the articles by the user, so that the analysis of the similarity of the articles is one-sided, the recommendation accuracy is low, and the collaborative filtering recommendation algorithm has the problems of cold start and sparsity because the articles on the network platform are numerous and the behavior data of the articles generated by the user on the network platform is less.
Disclosure of Invention
The embodiment of the application provides an article recommendation method, an article recommendation device and a storage medium, and can solve the problems that in the related art, the analysis of article similarity is one-sided, and the recommendation accuracy is low. The technical scheme is as follows:
in a first aspect, an item recommendation method is provided, the method including:
determining a first prediction score of a plurality of articles in an article set through a collaborative filtering model according to article behavior data of a target user, wherein the prediction score of each article is used for indicating the interest degree of the target user in each article;
determining k first items to be recommended to the target user from the plurality of items according to the first prediction scores of the plurality of items, wherein k is a positive integer;
determining first entity vectors corresponding to the k first articles respectively according to the k first articles and a knowledge graph vector set, wherein the knowledge graph vector set comprises first entity vectors of all entities in a knowledge graph, and the knowledge graph comprises a plurality of entities corresponding to the articles one by one and entity relations between every two entities in the entities;
determining second prediction scores of the k first items according to first entity vectors corresponding to the k first items respectively and item behavior data of the target user;
recommending the k first items to the target user if the mean square error between the first prediction scores of the k first items and the second prediction scores of the k first items is less than or equal to an error threshold.
Optionally, the determining second prediction scores of the k first items according to the first entity vectors corresponding to the k first items respectively and the item behavior data of the target user includes:
determining the similarity of every two items in the k first items according to the similarity between the first entity vectors corresponding to every two items in the k first items;
determining an article similarity matrix of the k first articles according to the similarity of every two articles in the k first articles, wherein the article similarity matrix of the k first articles comprises the similarity of every two articles in the k first articles;
and determining second prediction scores of the k first items according to the item similarity matrix of the k first items and the item behavior data of the target user.
Optionally, the method further comprises:
if the mean square error between the first prediction scores of the k first items and the second prediction scores of the k first items is larger than the error threshold, updating the item collaborative filtering model according to the second prediction scores of the k first items;
determining a third prediction score of the plurality of articles through the updated article collaborative filtering model according to the article behavior data of the target user;
determining k second items to be recommended from the plurality of items according to the third prediction scores of the plurality of items;
and recommending the object to the target user according to the k second objects.
Optionally, the recommending the target user for the item according to the k second items includes:
determining first entity vectors corresponding to the k second articles according to the k second articles and the knowledge graph vector set;
determining fourth prediction scores of the k second articles according to the first entity vectors corresponding to the k second articles respectively and the article behavior data of the target user;
recommending the k second items to the target user if the mean square error between the third prediction scores of the k second items and the fourth prediction scores of the k second items is less than or equal to the error threshold.
Optionally, the knowledge-graph includes a plurality of triples, each triplet including a head entity, a tail entity, and an entity relationship between the head entity and the tail entity, where the head entity and the tail entity are entities corresponding to any two of the plurality of articles;
before determining the first entity vectors corresponding to the k first items respectively according to the k first items and the knowledge graph vector set, the method further includes:
determining vector representations of all entities in a plurality of triples in an entity space and vector representations of all entity relationships in the triples in a relationship space, which are included in the knowledge graph, to obtain initial triplet vectors of the triples, where the initial triplet vector of each triplet includes a second entity vector of a head entity, a second entity vector of a tail entity, and a second relationship vector of an entity relationship in each triplet;
training the initial triplet vectors of the multiple triplets through a translation model to obtain target triplet vectors of the multiple triplets, wherein the target triplet vector of each triplet comprises a first entity vector of a head entity, a first entity vector of a tail entity and a first relation vector of an entity relation in each triplet, the distance between a reference vector in each triplet and the first entity vector of the tail entity is smaller than a distance threshold, and the reference vector is the vector sum of the first entity vector of the head entity and the first relation vector of the entity relation.
Optionally, the method further comprises:
in the process of training the initial triple vectors of the triples through the translation model, if a similar triple set exists in the triples, optimizing at least one triple in the similar triple set to obtain at least one optimized triple, and determining the initial triple vector of the at least one optimized triple; the initial triple vectors of all triples in the similar triple set are the same, the head entity and entity relations in all triples are correspondingly the same, and the difference degree between tail entities is greater than a difference degree threshold value;
and training the initial triple vector of the at least one optimized triple through the translation model to obtain a target triple vector of the at least one optimized triple.
Optionally, the optimizing at least one triplet in the similar triplet set to obtain at least one optimized triplet, and determining an initial triplet vector of the at least one optimized triplet includes:
re-determining an entity relationship between a head entity and a tail entity in the reference triple to obtain an optimized entity relationship, wherein the reference triple is any triple in the at least one triple;
determining the optimized entity relationship and a head entity and a tail entity in the reference triple as an optimized triple corresponding to the reference triple;
performing feature extraction on the optimized entity relationship to obtain an optimized relationship vector of the optimized entity relationship;
and determining the optimization relationship vector of the optimization entity relationship, the second entity vector of the head entity and the second entity vector of the tail entity in the reference triple as the initial triple vector of the optimization triple.
Optionally, the performing feature extraction on the optimized entity relationship to obtain an optimized relationship vector of the optimized entity relationship includes:
obtaining description information corresponding to the optimized entity relationship;
determining text keywords in the description information;
determining word vectors of the text keywords through a word vector model according to the text keywords, wherein the word vector model is used for determining the word vectors of any participle;
performing convolution processing on the word vectors of the text keywords to obtain a convolution result;
and performing pooling treatment on the convolution result to obtain the optimization relation vector.
In a second aspect, there is provided an item recommendation apparatus, the apparatus comprising:
the first determining module is used for determining first prediction scores of a plurality of articles in an article set through a collaborative filtering model according to article behavior data of a target user, wherein the first prediction score of each article is used for indicating the interest degree of the target user in each article;
a second determining module, configured to determine, according to the first prediction scores of the multiple items, k first items to be recommended to the target user from the multiple items, where k is a positive integer;
a third determining module, configured to determine, according to the k first items and a knowledge-graph vector set, first entity vectors corresponding to the k first items, respectively, where the knowledge-graph vector set includes first entity vectors of entities in a knowledge graph, and the knowledge graph includes a plurality of entities in one-to-one correspondence with the plurality of items and entity relationships between every two entities in the plurality of entities;
a fourth determining module, configured to determine second prediction scores of the k first items according to the first entity vectors corresponding to the k first items, respectively, and the item behavior data of the target user;
a first recommending module, configured to recommend the k first items to the target user if a mean square error between the first prediction scores of the k first items and the second prediction scores of the k first items is less than or equal to an error threshold.
Optionally, the fourth determining module includes:
the first determining submodule is used for determining the similarity of every two items in the k first items according to the similarity between the first entity vectors corresponding to every two items in the k first items;
the second determining submodule is used for determining an article similarity matrix of the k first articles according to the similarity of every two articles in the k first articles, wherein the article similarity matrix of the k first articles comprises the similarity of every two articles in the k first articles;
and the third determining submodule is used for determining second prediction scores of the k first items according to the item similarity matrix of the k first items and the item behavior data of the target user.
Optionally, the apparatus further comprises:
an updating module, configured to update the collaborative filtering model according to the second prediction scores of the k first items if a mean square error between the first prediction scores of the k first items and the second prediction scores of the k first items is greater than the error threshold;
a fifth determining module, configured to determine, according to the item behavior data of the target user, a third prediction score of the multiple items through an updated collaborative filtering model;
a sixth determining module, configured to determine k second items to be recommended from the plurality of items according to third prediction scores of the plurality of items;
and the second recommending module is used for recommending the object to the target user according to the k second objects.
Optionally, the second recommending module includes:
a fourth determining submodule, configured to determine, according to the k second products and the knowledge-graph vector set, first entity vectors corresponding to the k second products, respectively;
a fifth determining sub-module, configured to determine fourth prediction scores of the k second items according to the first entity vectors corresponding to the k second items, respectively, and the item behavior data of the target user;
and the recommending submodule is used for recommending the k second items to the target user if the mean square error between the third prediction scores of the k second items and the fourth prediction scores of the k second items is less than or equal to the error threshold.
Optionally, the knowledge graph includes a plurality of triples, each triplet including a head entity, a tail entity, and an entity relationship between the head entity and the tail entity, where the head entity and the tail entity are entities corresponding to any two of the plurality of items;
the device, still include:
a seventh determining module, configured to determine a vector representation of each entity in the multiple triples in the knowledge graph in an entity space and a vector representation of each entity relationship in the multiple triples in a relationship space, to obtain an initial triplet vector of the multiple triples, where the initial triplet vector of each triplet includes the second entity vector of the head entity, the second entity vector of the tail entity, and the second relationship vector of the entity relationship in each triplet;
the first training module is configured to train the initial triplet vectors of the multiple triplets through a translation model to obtain target triplet vectors of the multiple triplets, where the target triplet vector of each triplet includes a first entity vector of a head entity, a first entity vector of a tail entity, and a first relationship vector of an entity relationship in each triplet, and a distance between a reference vector in each triplet and the first entity vector of the tail entity is smaller than a distance threshold, and the reference vector is a vector sum of the first entity vector of the head entity and the first relationship vector of the entity relationship in each triplet.
Optionally, the apparatus further comprises:
the optimization module is configured to, in a process of training the initial triplet vectors of the multiple triples through the translation model, if a similar triplet set exists in the multiple triples, optimize at least one triplet in the similar triplet set to obtain at least one optimized triplet, and determine the initial triplet vector of the at least one optimized triplet; the initial triple vectors of all triples in the similar triple set are the same, the head entity and entity relations in all triples are correspondingly the same, and the difference degree between tail entities is greater than a difference degree threshold value;
and the second training module is used for training the initial triple vector of the at least one optimized triple through the translation model to obtain a target triple vector of the at least one optimized triple.
Optionally, the optimization module includes:
a sixth determining sub-module, configured to re-determine, for a reference triple in the at least one triple, an entity relationship between a head entity and a tail entity in the reference triple to obtain an optimized entity relationship, where the reference triple is any triple in the at least one triple;
a seventh determining sub-module, configured to determine the optimized entity relationship, and a head entity and a tail entity in the reference triple as an optimized triple corresponding to the reference triple;
the feature extraction submodule is used for extracting features of the optimized entity relationship to obtain an optimized relationship vector of the optimized entity relationship;
and the eighth determining submodule is used for determining the optimization relationship vector of the optimization entity relationship, the second entity vector of the head entity and the second entity vector of the tail entity in the reference triple as the initial triple vector of the optimization triple.
Optionally, the feature extraction sub-module includes:
the obtaining subunit is used for obtaining the description information corresponding to the optimized entity relationship;
the first determining subunit is used for determining the text keywords in the description information;
the second determining subunit is used for determining word vectors of the text keywords through a word vector model according to the text keywords, and the word vector model is used for determining the word vectors of any participle;
the convolution submodule is used for performing convolution processing on the word vectors of the text keywords to obtain a convolution result;
and the pooling submodule is used for pooling the convolution result to obtain the optimization relation vector.
In a third aspect, an article recommendation device is provided, which is characterized by comprising:
a processor;
a memory for storing processor-executable instructions;
wherein the processor is configured to perform the steps of any of the methods of the first aspect.
In a fourth aspect, a computer-readable storage medium is provided, the computer-readable storage medium having stored thereon instructions, wherein the instructions, when executed by a processor, implement the steps of any one of the methods of the first aspect.
In a fifth aspect, there is provided a computer program product comprising instructions which, when run on a computer, cause the computer to perform the steps of any of the methods of the first aspect described above.
The technical scheme provided by the embodiment of the application has the following beneficial effects:
in the embodiment of the application, first prediction scores of a plurality of articles are determined through a collaborative filtering model according to article behavior data of a target user, k first articles to be recommended are determined according to the first prediction scores of the plurality of articles, then first entity vectors of the k first articles are determined according to the k first articles and a knowledge graph vector set, second prediction scores of the k first articles are determined according to the first entity vectors of the k first articles and the article behavior data of the target user, and if a mean square error between the first prediction scores and the second prediction scores of the k first articles is smaller than or equal to an error threshold, the k first articles are recommended to the target user. That is, when recommending articles through the collaborative filtering recommendation model, the entity vector of the article in the knowledge graph can be used as auxiliary information for recommendation, so that on the basis of determining the similarity of the articles according to the article behavior data, the similarity between the article attributes is fully considered, the article is accurately predicted and scored by combining the article behavior data of the user and the self-attributes of the article, the recommendation accuracy can be improved, and the problems of cold start and sparseness existing in the collaborative filtering recommendation algorithm are avoided.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
Fig. 1 is a flowchart of a training method of an item recommendation system according to an embodiment of the present application;
FIG. 2 is a schematic diagram of a cross-learning item recommendation system provided by an embodiment of the present application;
FIG. 3 is a flowchart of an item recommendation method provided by an embodiment of the present application;
FIG. 4 is a diagram illustrating a one-to-many relationship in a triplet provided by an embodiment of the present application;
FIG. 5 is a schematic diagram of a convolution process provided by an embodiment of the present application;
FIG. 6 is a schematic diagram of a pooling process provided by an embodiment of the present application;
FIG. 7 is a schematic diagram of another convolution process provided by embodiments of the present application;
FIG. 8 is a schematic diagram of another pooling process provided by embodiments of the present application;
FIG. 9 is a model diagram of an improved translation model provided by an embodiment of the present application;
FIG. 10 is a schematic structural diagram of an article recommendation device according to an embodiment of the present application;
fig. 11 is a schematic structural diagram of an electronic device according to an embodiment of the present application.
Detailed Description
To make the objects, technical solutions and advantages of the present application more clear, embodiments of the present application will be described in further detail below with reference to the accompanying drawings.
In the embodiment of the application, in order to solve the problems of one-sided object similarity analysis, low recommendation accuracy and cold start and sparsity of a collaborative filtering recommendation algorithm in a collaborative filtering recommendation system, the object recommendation method fusing the collaborative filtering recommendation algorithm based on objects and a knowledge graph technology is provided. Before describing the item recommendation method provided in the embodiment of the present application in detail, a model training process in the embodiment of the present application is described in detail.
Fig. 1 is a flowchart of a training method of an item recommendation system according to an embodiment of the present application, where the method is applied to an electronic device, and the electronic device may be an electronic device such as a terminal or a server, which is not limited in this embodiment of the present application, and as shown in fig. 1, the method includes the following steps:
step 101: and acquiring initial triple vectors of a plurality of triples in the knowledge graph, and training the initial triple vectors of the triples through a translation model to obtain a vector training result.
It should be noted that the knowledge graph includes an entity corresponding to an actual article and an entity relationship between two entities, and the entity is obtained by mapping the article to a knowledge graph space. Further, one entity of any two entities in the knowledge-graph may be referred to as a head entity, the other entity as a tail entity, and the head entity, the tail entity, and the entity relationship of the head entity and the tail entity as a triplet. That is, the knowledge-graph includes a plurality of triples, each triplet including a head entity, a tail entity, and an entity relationship of the head entity and the tail entity.
In addition, the initial triplet vector for each triplet includes the entity vector for the head entity, the entity vector for the tail entity, and the relationship vector for the entity relationships in each triplet. For convenience of explanation, the entity vector of the head entity may also be referred to as a head entity vector, and the entity vector of the tail entity may also be referred to as a tail entity vector. An entity vector in the initial triplet vectors of the plurality of triples is a vector representation of each entity in an entity space, and a relationship vector in the initial triplet vectors is a vector representation of each entity in a relationship space. As an example, the entity vector may be obtained by performing feature extraction on an entity, and the relationship vector may be obtained by performing feature extraction on an entity relationship.
Furthermore, the vector training result includes a plurality of target triplet vectors of triples, and a distance between a reference vector in the target triplet vectors and an entity vector of a tail entity is smaller than a distance threshold, and the reference vector of the target triplet vectors refers to a vector sum of an entity vector of a head entity of the target triplet vectors and a relationship vector of an entity relationship. Wherein the distance threshold may be preset.
It should be noted that the initial triplet vectors of each triplet in the knowledge graph may not accurately indicate a complex semantic relationship between an entity and a relationship, and therefore in the embodiment of the present application, the initial triplet vectors of the multiple triplets need to be trained through a translation model, so that the trained triplet vectors can accurately indicate the complex semantic relationship between the entity and the relationship between the entities.
The translation model is used to model entities and entity relationships to embed the entities and entity relationships into a low-dimensional vector space. That is, the translation model is used to embed the entities and entity relationships in the input initial triplet vectors into the low-dimensional vector space. Illustratively, the above translation model may be a Trans E model (a kind of translation model).
Additionally, the translation model may view entity relationships in the knowledge-graph as translation vectors between entities. For example, assume a triplet is (h, r, t), where h is a head entity, t is a tail entity, r is an entity relationship between the head entity and the tail entity, and a triplet vector corresponding to the triplet (h, r, t) is (h, r, t)
Figure BDA0002427308760000101
Wherein +>
Figure BDA0002427308760000102
For a head entity vector, <' > based on>
Figure BDA0002427308760000103
Is a vector of the tail entity, and is,
Figure BDA0002427308760000104
is a relationship vector, then for each initial triplet vector (h, r, t), the relationship vector @maybe used>
Figure BDA0002427308760000105
As head entity vector->
Figure BDA0002427308760000106
And a trailing entity vector>
Figure BDA0002427308760000107
Can also be moved between->
Figure BDA0002427308760000108
Regarded as->
Figure BDA0002427308760000109
To>
Figure BDA00024273087600001010
The translation of (2).
The translation model considers that the triplet vector of a correct triplet should satisfy the vector between the head entity vector and the relational entity vector and should be equal to the tail entity vector. The translation model considers a correct (h, r, t) triplet vector
Figure BDA00024273087600001011
Shall satisfy >>
Figure BDA00024273087600001012
Similarly, if an error triple exists, the direction of the triple does not satisfy the relationship. Thus, the translation model may define a vector distance to represent the distance between a reference vector, which is the vector sum of the entity vector referring to the head entity and the relationship vector of the entity relationship, and the entity vector of the tail entity. The smaller the vector distance, the more likely the triplet is to be a correct triplet, and the larger the vector distance, the less likely the triplet is to be a correct triplet. For example, the vector distance may be ≦ ≦>
Figure BDA00024273087600001013
Indicates that for a triplet (h, r, t), then based on the number of bins, the value of a triplet is greater than the value of a triplet>
Figure BDA00024273087600001014
The smaller the distance, the better.
That is, the goal of the translation model is to make the distance between the reference vector of the triplet vector of each triplet in the knowledge-graph and the entity vector of the tail entity as small as possible, i.e., to make
Figure BDA00024273087600001015
The smaller the distance, the better. As an example, the objective function of the translation model is shown in the following equation (1):
Figure BDA0002427308760000111
where S is a set of positive sample triples, S ' is a set of negative sample triples, γ represents the spacing between positive and negative samples, [ γ + d (h + r, t) -d (h ' + r, t ')] + Represents max [0, γ + d (h + r, t) -d (h '+ r, t')]That is, a binary task is performed on a given triplet to distinguish whether it is a positive sample or a negative sample, for example, an initial triplet that meets the objective function is determined as a target triplet vector, and a further adjustment operation is performed on an initial triplet that does not meet the objective function. For example, negative sample triplets may be constructed by themselves by substitution, with the goal of maximizing the nearest positive and negative sample distance.
In a possible implementation manner, the translation model embeds an entity and a relationship in an input initial triplet vector into a low-dimensional vector space, judges whether the initial triplet vector meets a distance condition, takes the initial triplet vector meeting the distance condition as a target triplet vector, re-determines the relationship between a head entity and a tail entity for the initial triplet vector not meeting the distance condition, and further adjusts the initial triplet vector so that the initial triplet vector can meet the distance condition. That is, the initial triplet vectors of the triples are trained through the translation model, and the entities and the relationships can be embedded into the low-dimensional vector space to obtain a target triplet vector (h + r) which is as equal to t as possible. Wherein the distance condition may be that a distance between the reference vector and the entity vector of the tail entity is less than a distance threshold.
As an example, when obtaining the initial triplet vectors of the multiple triplets from the knowledge graph, multiple user IDs (identities) of all users may be determined, the multiple triplets in the knowledge graph are obtained through different APIs (Application Programming interfaces) in the graph data according to the multiple user IDs, the initial triplet vector of each triplet is obtained, and the initial triplet vectors of the multiple triplets are input to the translation model as one input set.
As an example, in order to accurately obtain recommendation data related to the industry and further recommend an item more conforming to the user interest to the user, before step 101 is executed, a data dictionary configuration of the item and the map entity of the user to be recommended needs to be established according to the existing knowledge map. It should be noted that, the embodiments of the present application do not relate to a specific knowledge graph construction technology, and a knowledge graph capable of being used for learning and training of a translation model can be obtained by querying based on an existing knowledge graph database by using a mapping rule and a graph database.
The knowledge maps of different industries deposit different data, the recommended data sets comprise users, articles and explicit feedback and implicit feedback of the users to the articles, and the data sets can be cleaned to deposit a low-sparsity user set, an article set and a feedback set related to the articles of the users. In the subsequent online article recommendation process, in order to acquire recommendation data related to the industry in the existing knowledge graph, namely the knowledge graph corresponding to the data set deposited after the data is cleaned, data dictionary configuration of the article and the graph entity of the user to be recommended needs to be established.
As one example, the data set may be cleaned using an ETL (Extract-Transform-Load) cleaning technique. The ETL cleaning is used for describing the process of extracting, interactively converting and loading data from a source end to a destination end, and aims to screen and filter the data which do not meet the requirements and deposit and store the required data in a database, so that the follow-up data query and call are facilitated. The displayed feedback is to represent the user's hobbies, usually some numbers or data, in some direct way, and may be, for example, the user's rating, etc. of the item. Implicit feedback represents the user's interests in some indirect way, such as user search data, user click-through data, user's status of items that have entered the shopping cart, etc.
In addition, the data dictionary configuration for establishing the user item to be recommended and the map entity is realized in many ways, for example, if the knowledge map supports a visual interface display, a custom template configuration way can be selected. Or, a recommendation template component mode can be selected for configuration, and the recommendation template component interface can be used for establishing a mapping set of the map entity corresponding to the user article and the map relation corresponding to the associated feedback for the recommendation data set in an automatic configuration mode and a manual editing and correcting mode by butting the settled recommendation data set. And finally, storing the result set indexes into the graph data so as to carry out graph query operation.
That is, in the graph data, the user or article ID already has a corresponding entity ID in the knowledge graph, the user ID of the user is input through graph retrieval, and the history data of each user is acquired through an API interface. And then, the user ID of the user can be input, and the historical data of the user can be obtained from the API interface corresponding to the user ID according to the corresponding relation between the user ID and the API interface. The historical data may be explicit feedback and implicit feedback of the user to the item.
In the manner of graph query described above, a plurality of triples may be obtained in the knowledge-graph of graph data, and these triples will be used as an input set of graph embedding learning training. In addition, for the knowledge graph determined as above, with the help of the existing word vector model, the initial triple vector corresponding to each triple in the knowledge graph can be determined, and all the initial triple vectors as well as the knowledge graph are stored in the graph data.
Step 102: and determining first prediction scores of a plurality of items in the item set through a collaborative filtering model according to the item behavior data of the sample user, wherein the first prediction score of each item is used for indicating the interest degree of the sample user for each item.
It should be noted that the sample user refers to a user who satisfies the model training requirement. For example, the sample user may be any registered user in the related item recommendation system. In order to ensure the learning of the characteristics of the knowledge map and the training effect of the collaborative filtering model, registered users with rich display feedback and implicit feedback can be screened from the registered users, and the rich article behavior data of the users can provide better training samples for the collaborative filtering model so as to detect the training effect of the collaborative filtering model.
Still further, the item behavior data of the sample user may include item browsing lists, item purchasing lists of the sample user, behavior data of the user such as rating, praise, collection, and the like of the items.
In addition, the collaborative filtering model may be an article-based collaborative filtering model, in the embodiment of the present application, distance description is performed on the article-based collaborative filtering model, and other collaborative filtering models may also be used, which is not limited in the present application.
The calculation process of the collaborative filtering model based on the articles mainly comprises the following two processes of calculating the similarity between the articles and recommending and predicting:
1) Calculating similarity between items
Determining similarity matrixes of a plurality of articles in the article set according to article behavior data of a plurality of users, wherein the similarity matrixes of the plurality of articles comprise the similarity between every two articles in the plurality of articles.
Illustratively, determining each similarity in the similarity matrix may be achieved by the following equation (2):
Figure BDA0002427308760000131
wherein, ω is ij Is an element in the similarity matrix for indicating the similarity between item i and item j, | N (i) | is the number of users scoring item i, | N (j) | is the number of users scoring item j, and N (i) andnun (j) | is the number of users scoring item i and item j simultaneously. Therefore, the above formula (2) can determine how many proportion of users who like the item i also like the item j, and can further determine the similarity between the item i and the item j according to the proportion of the users who like the item i and the item j at the same time.
Further, the similarity matrix of a plurality of articles in the article set can be determined according to the article behavior data and the similarity weights of a plurality of users. The similarity distance is used to avoid over-biasing the value of the item similarity.
Illustratively, after introducing the similarity weight, the similarity matrix may be determined by the following formula (3):
Figure BDA0002427308760000132
wherein β is a shrinkage coefficient, and when | N (i) # N (j) | is small, it will play a role of shrinkage, ensuring that item recommendation can be performed for sample users with less item behavior data.
If item j is very hot, the more users there are item behavior data for item j, then ω is determined in equation (2) ij It will be large, even close to 1. Therefore, when recommending articles for sample users, in order to avoid recommending popular articles, important measures of similarity weight can be added in the determination of the similarity matrix of the articles to avoid the problem that the determined similarity values are excessively biased.
2) Recommendation prediction
A first prediction score for each of a plurality of items in the set of items is determined based on the similarity matrix for the plurality of items and the item behavior data for the sample user. The first prediction score is used to indicate how much the sample user likes the item. The first prediction score may be a percentage system or a tenth system, which is not limited in this application.
In one possible implementation, after the item-item similarity matrix is obtained, the item-based collaborative filtering recommendation algorithm may perform predictive scoring using the following equation (4):
Figure BDA0002427308760000141
wherein, P uj Representing the interest degree of the sample user u in the item j, namely a first prediction score of the item j; n (u) represents a sample user u's item behavior data set; i is an item for which the sample user u has item behavior data, e.g., i is an item that the sample user has scored; s (j, k) represents k articles with highest similarity to the article iJ is an item in the set, ω ji Representing the degree of similarity, r, between item j and item i ui The like degree of the sample user for the item i is shown, and the like degree can be determined according to any behavior data in the item behavior data of the sample user for the item i, which is not limited in the embodiment of the present application.
Step 103: k items to be recommended to the sample user are determined from the plurality of items according to the first prediction scores for the plurality of items.
Wherein k may be preset, for example, k is 5 or 10, etc.
In one possible implementation, the top k items are determined from the plurality of items as the k items to be recommended to the sample user in descending order of the first prediction score.
In another possible implementation manner, an item with a first prediction score larger than a score threshold value may be selected from a plurality of items, and the selected item may be used as k items to be recommended to the sample user.
Step 104: and acquiring entity vectors of the k articles from the vector training result according to the k articles.
In a possible implementation manner, the corresponding entity ID may be searched in the knowledge graph according to the article ID, and then the entity vectors of k articles are obtained in the vector training result according to the entity ID.
It should be noted that the entity vectors of k items at this time satisfy the relationship between the entity vectors, that is, the entity vectors of k items at this time can well reflect the semantic correlation between the items.
Step 105: and determining the similarity matrix of the k articles according to the entity vectors of the k articles.
In a possible implementation manner, after the entity vectors of the k items are determined, the similarity of every two items in the k items can be determined based on the similarity between the entity vectors corresponding to every two items in the k items, and then the similarity matrix corresponding to the k items is determined according to the similarity of every two items in the k items. That is, the similarity between the entity vectors of any two of the k items may be taken as the similarity between the two items.
Step 106: and determining second prediction scores of the k items according to the similarity matrix of the k items and the item behavior data of the sample user.
For example, the second prediction scores for the k items may be determined by a collaborative filtering model based on the similarity matrix for the k items and the item behavior data of the sample user.
For example, the first prediction scores of each of the multiple items may be determined according to the similarity matrix of the multiple items and the item behavior data of the sample user, and the second prediction scores of the k items may be determined according to the similarity matrix of the k items and the item behavior data of the sample user, which is not described herein again in this embodiment of the present application.
Illustratively, the similarity between item a and item B is 70%, in the case of the ten-degree system, the score of the sample user of item a is 7, the score of the sample user of item B is 2, and since the similarity between item a and item B is 70% under the consideration of adding the semantic vector, the second prediction score for item B is also adjusted. For example, 4.9 may be adjusted, where 4.9 is determined based on 7 × 70% = 4.9.
Step 107: a mean square error between the first prediction scores for the k items and the second prediction scores for the k items is determined.
Wherein the first prediction score is determined based on the collaborative filtering model, the second prediction score is determined based on the similarity between the entity vectors in the vector training result in the translation model and the item behavior data of the sample user, and for any item in the k items, the first prediction score of the item may be the same as or different from the second prediction score. The items recommended by the collaborative filtering model with the same prediction scores are consistent with the interests of the sample user in semantic relevance, the items recommended by the collaborative filtering model with different prediction scores are only recommended to the user structurally, and the recommended items do not completely meet the semantic relevance.
Step 108: and if the mean square error is larger than the error threshold, starting the cross learning task.
The cross learning task may be a task of continuing to train the collaborative filtering model according to the second prediction scores of the k items, or may also be a task of continuing to train the triple vector through the translation model, which is not limited in the embodiment of the present application. In addition, the cross learning task may also be a task of training the collaborative filtering model and the triplet vectors alternately, which is not limited in the embodiment of the present application.
That is, in the embodiment of the present application, after the cross learning task is started, parameters of the collaborative filtering model may be fixed, model parameters of the translation model may be trained, model parameters of the translation model may also be fixed, and parameters of the collaborative filtering model may be trained, or of course, the two may also be trained alternately until convergence.
Step 109: and if the mean square error is less than or equal to the error threshold, ending the cross learning task.
If the mean square error is less than or equal to the error threshold, it indicates that the cross training process reaches convergence, and the task can be ended at this time. Through cross training of the collaborative filtering model and the triple vectors, the recommendation result of auxiliary recommendation according to the learning result can be more accurate.
For example, the training process of the item recommendation system may be as shown in fig. 2, and as shown in fig. 2, the item recommendation system may be a cross-learning item recommendation system, if the mean square error is greater than the error threshold, a cross-learning task switch may be turned on, and a cross-learning task is started, and if the mean square error is less than the error threshold, the task is ended.
Next, a detailed description will be given of an item recommendation method provided in an embodiment of the present application,
fig. 3 is a flowchart of an article recommendation method provided in an embodiment of the present application, where the method is applied to an electronic device, where the electronic device may be an electronic device such as a terminal or a server, as shown in fig. 3, the method includes the following steps:
step 301: and determining first prediction scores of a plurality of items in the item set through a collaborative filtering model according to the item behavior data of the target user, wherein the first prediction score of each item is used for indicating the interest degree of the target user in each item.
It should be noted that the target user is a user who is to make item recommendation. When article recommendation needs to be performed on a target user, article behavior data of the target user can be obtained, and then first prediction scores of a plurality of articles in an article set are determined through a collaborative filtering model according to the article behavior data of the target user. For example, the target user may be an old user registered in the item recommendation system, or may be a new user of the item recommendation system, which is not limited in this embodiment of the application.
In a possible implementation manner, the first prediction scores of the multiple articles in the article set may be determined through the collaborative filtering model according to the article behavior data of the sample user in step 102, and the first prediction scores of the multiple articles in the article set may be determined through the collaborative filtering model according to the article behavior data of the target user, which is not described herein again in detail, and see step 102 for details.
The item set is a set of items to be recommended to a target user. As one example, for an item that the target user has never focused on, the first prediction score for the item may be 0. For example, the object set includes 1000 objects of four categories of 200 articles for daily use, 100 snacks, 300 stationeries and 400 clothes, but the object behavior data of the target user is only stationery-related object behavior data, and in this case, the corresponding first prediction score is determined for only 300 stationeries, and the first prediction score related to articles for daily use, snacks and clothes can be 0. For another example, when a piece of apparel is printed with a stationery pattern, the apparel may also be added to the user's collection of items and a first prediction score for the apparel may be determined.
Step 302: and determining k first items to be recommended to the target user from the plurality of items according to the first prediction scores of the plurality of items, wherein k is a positive integer.
In one possible implementation, the top k items from the plurality of items may be determined as the k first items to be recommended to the sample user in order of the first prediction scores from large to small.
In another possible implementation manner, an item with a first prediction score larger than a score threshold value may be selected from a plurality of items, and the selected item is used as k first items to be recommended to the sample user.
In an embodiment, the prediction scores of some items in the item set may be 0, in order to improve the accuracy of recommendation for the user in the embodiment of the present application, screening may be performed in the item set first, to determine a plurality of items of which the first prediction scores are not 0, and then determine, according to the above manner, K first items to be recommended to the target user from the plurality of items.
Step 303: determining first entity vectors corresponding to the k first articles respectively according to the k first articles and a knowledge graph vector set, wherein the knowledge graph vector set comprises first entity vectors of all entities in a knowledge graph, and the knowledge graph comprises a plurality of entities corresponding to the articles one by one and entity relations between every two of the entities.
It should be noted that the knowledge graph may include a plurality of triples, each triplet includes a head entity, a tail entity, and an entity relationship between the head entity and the tail entity, and the head entity and the tail entity may be entities corresponding to any two articles in the plurality of articles.
In addition, based on the knowledge graph, in order to accurately reflect the description of the relationship between the entities in the knowledge graph, the entities and the vectors in the knowledge graph can be embedded into a low-dimensional vector to determine an initial triple vector corresponding to each triple in a plurality of triples in the knowledge graph, the initial triples of the triples are trained through a translation model to obtain target triple vectors of the triples, and a knowledge graph vector set is constructed by the target triple vectors of the triples.
The target triple vectors of the triples can be obtained by training the initial triple vectors of the triples through a translation model. And the initial triple vector of each triple comprises a second entity vector of the head entity, a second entity vector of the tail entity and a second relation vector of the entity relation in each triple. The initial triplet vector for the plurality of triples may be derived by determining a vector representation in entity space for each entity in the plurality of triples and a vector representation in relationship space for each entity in the plurality of triples.
The set of knowledge-graph vectors may include a target triplet vector for a plurality of triples, the target triplet vector for each triplet including the first entity vector for the head entity, the first entity vector for the tail entity, and the first relationship vector for the entity relationship in each triplet, and a distance between the reference vector in each triplet and the first entity vector for the tail entity being less than a distance threshold, the reference vector for each triplet referring to a vector sum of the first entity vector for the head entity and the first relationship vector for the entity relationship in each triplet.
Since translation models are generally unable to handle complex relationships, when the relationships are one-to-many or many-to-many, entity representations learned by the translation models may be less discriminative. In order to solve the problem, in the embodiment of the present application, in the process of training the initial triplet vectors of the multiple triplets through the translation model, a similar triplet set may be further determined from the multiple triplets, at least one triplet in the similar triplet set is optimized to obtain at least one optimized triplet, the initial triplet vector of the at least one optimized triplet is determined, and then the initial triplet vector of the at least one optimized triplet is trained through the translation model to obtain the target triplet vector of the at least one optimized triplet.
The initial triple vectors of all the triples in the similar triple set are the same, the head entity and the entity relationship in all the triples are correspondingly the same, and the difference degree between the tail entities is greater than the difference degree threshold value. That is, a similar triplet refers to a triplet vector
Figure BDA0002427308760000181
Head entity h and entity relationship r are the same, butThere are triples of large differences between tail entities t.
And calculating the difference between the tail entities according to the attribute information of the tail entities. For example, the similarity between the attribute information of two tail entities is calculated, and if the similarity between the attribute information of two tail entities is smaller than a similarity threshold, it is determined that the difference between the two tail entities is greater than a difference threshold. As an example, the similarity between the attribute information of the tail entities may be determined by a Simhash algorithm.
As an example, the penalty function for a translation model may be as shown in equation (5) below:
Figure BDA0002427308760000182
wherein L is a loss function; δ =1 or δ =0; s is a set of positive sample triples, S' is a set of negative sample triples, γ represents the distance between positive and negative samples, [ x ]] + Represents max [0, x ]]。
As can be seen from equation (5), if the similarity between the attribute information of two tail entities is smaller than the similarity threshold, the two tail entities can be considered to be the same. For example, if two tail entities are r respectively 1 And r 2 If r is 1 And r 2 Is less than the similarity threshold, then r can be considered to be at that time 1 ≈r 2 。
As an example, optimizing at least one triplet of the set of similar triples to obtain at least one optimized triplet, and determining an initial triplet vector for the at least one optimized triplet includes: re-determining an entity relationship between a head entity and a tail entity in the reference triple to obtain an optimized entity relationship for the reference triple in the at least one triple, wherein the reference triple is any triple in the at least one triple; then determining the optimized entity relationship and the head entity and the tail entity in the reference triple as the optimized triple corresponding to the reference triple; performing feature extraction on the optimized entity relationship to obtain an optimized relationship vector of the optimized entity relationship; and determining the optimized relation vector of the optimized entity relation as an initial triplet vector of the optimized triplet by referring to the second entity vector of the head entity and the second entity vector of the tail entity in the triplet.
That is, in the embodiment of the present application, an entity with similar tail entities may be considered as one entity, the tail entities may be obtained by training according to a default r, if the difference between the tail entities is large, the tail entities may not be the same entity, and the relationship vector of the entity relationship is recalculated, and then a new tail entity is obtained by training.
For example, when the head entity is "fruit" and the relationship is "apple genus", the determined tail entity is "red apple" and "green apple", but the similarity between the attribute information of the tail entity "red apple" and "green apple" is greater than the similarity threshold, the two triplets of (fruit, apple, red apple) and (fruit, apple, green apple) are determined as similar triplets, the triplets (fruit, apple, red apple) can be determined as reference triplets, and the entity relationship between the head entity "fruit" and the tail entity "red apple" in the reference triplets can be re-determined, so that the optimized entity relationship "red apple genus" can be obtained, the optimized entity relationship, and the head entity and the tail entity in the reference triplets are determined as optimized triplets corresponding to the reference triplets, where the optimized triplets are (fruit, red apple). Similarly, the (fruit, apple and green apple) may also be determined as the reference triplet, and the entity relationship between the head entity "fruit" and the tail entity "green apple" in the reference triplet is re-determined, so as to obtain the optimized entity relationship "green apple genus", and the optimized entity relationship, and the head entity and the tail entity in the reference triplet are determined as the optimized triplet corresponding to the reference triplet, where the optimized triplet is the (fruit, green apple and green apple).
For another example, when the head entity is "fruit" and the relationship is "apple genus", the determined tail entity is "red apple" and "green apple", when it is determined that the similarity between the attribute information of the tail entity "red apple" and "green apple" is smaller than the similarity threshold, the tail entity is trained according to the default r to obtain the tail entity "red apple", and finally the determined triple tail (fruit, apple genus, red apple) is deleted. By way of example, there is a one-to-many relationship in a triple as shown in FIG. 4.
The extracting the features of the optimized entity relationship to obtain the optimized relationship vector of the optimized entity relationship may include: obtaining the description information corresponding to the optimized entity relationship; determining text keywords in the description information; determining a word vector of the text keyword through a word vector model according to the text keyword, wherein the word vector model is used for determining a word vector of any word segmentation; and extracting the characteristics of the word vector of the text keyword to obtain the optimized relation vector.
As an example, the optimized relationship vector may be obtained by performing feature extraction on the word vector of the text keyword through a feature extraction model.
The feature extraction model can be a CNN model, and feature extraction can be performed on the word vectors of the text keywords through the CNN model to obtain the optimized relation vector. It can also be said that the word vector of the text keyword is encoded through the CNN model to obtain the optimized relationship vector.
The CNN model may include convolutional layers and pooling layers, among others. Through the feature extraction model, when the feature extraction is carried out on the word vectors of the text keywords, the convolution processing can be carried out on the word vectors of the text keywords to obtain convolution results, and then the pooling processing is carried out on the convolution results to obtain optimized relation vectors.
Further, the CNN model may include two convolutional layers and two pooling layers, and the operation of extracting the feature of the word vector of the text keyword through the feature extraction model may include the following steps:
1) And performing convolution processing on the word vectors of the text keywords through the first convolution layer to obtain a first convolution result.
For example, each word vector may have dimensions of 100 dimensions. The convolution kernels of the first convolutional layer may include 100 convolution kernels, each convolution kernel being a 2 x 100 convolution kernel. As an example, n word vectors with the top positions may be selected from the word vectors of the text keyword, and then the first convolution layer may perform convolution processing on the n word vectors to obtain a first convolution result.
Referring to fig. 5, fig. 5 is a schematic diagram of a convolution process provided in an embodiment of the present application, and as shown in fig. 5, assuming that each word vector of a text keyword is 100 dimensions, 50 word vectors located at the front position may be selected from the word vectors of the text keyword as an input of a first convolution layer, and an input obtained is a matrix of 50 × 100, and then 100 convolution kernels CovA are obtained through the first convolution layer, and the size of each convolution kernel is set to be 2 × 100, and the input is convolved with 100 convolution kernels CovA, so that a first convolution result is obtained, and the first convolution result is 100 matrices with a size of 49 × 1.
2) And performing pooling treatment on the first convolution result through the first pooling layer to obtain a first pooling result.
For example, the pooling core of the first pooling layer may be 4 x 1 in size with a step size of 4 x 1. As an example, the first rolled result may be maximally pooled by a first pooling layer to obtain a first pooled result.
Referring to fig. 6, fig. 6 is a schematic diagram of a pooling process provided in an embodiment of the present application, and as shown in fig. 5, assuming that the input of the first pooling layer is 100 matrices with a size of 49 × 1, the pooled cores of the first pooling layer are 4 × 1 in size, and the step size is 4 × 1, the output of the first pooling layer is 100 matrices with 13 × 1.
3) And performing convolution processing on the first pooling result through the second convolution layer to obtain a second convolution result.
The convolution processing of the second convolutional layer is to have each position concatenated while aggregating across channels. For example, the second convolutional layer may take 100 convolutional kernels of size 1 x 100.
Referring to fig. 7, fig. 7 is a schematic diagram of another convolution process provided in the present embodiment, and as shown in fig. 6, assuming that the input of the second convolutional layer is 1 matrix of 13 × 100, 100 convolutional kernels with a size of 1 × 100 are selected for the second convolutional layer, and the output of the second convolutional layer is 100 matrices with a size of 13 × 1.
The input to the first pooling layer is 100 matrices of size 49 x 1, the pooled kernel of the first pooling layer is 4 x 1, and the output of the first pooling layer is 100 matrices of 13 x 1.
4) And performing pooling treatment on the second convolution result through a second pooling layer to obtain an optimized relation vector.
For example, the pooling cores of the second pooling layer may be 13 x 1 in size with a step size of 1 x 1. As an example, the second convolution result may be averaged and pooled by the second pooling layer to obtain the optimized relationship vector.
Referring to fig. 8, fig. 8 is a schematic diagram of another pooling process provided in the present embodiment, and as shown in fig. 8, assuming that the input of the second pooling layer is 100 matrices with a size of 13 × 1, the pooled kernel of the second pooling layer is 13 × 1, and the step size is 1 × 1, the output of the second pooling layer is 100 matrices with 1 × 1, so as to obtain a relationship vector representation with 100 dimensions.
In the embodiment of the application, in the aspect of knowledge graph representation learning, the existing translation model representation learning method is used as a basis, a new improved scheme is designed, and the similarity of the two aspects of the structure and the semantics is combined.
Referring to fig. 9, fig. 9 is a model diagram of an improved translation model provided in an embodiment of the present application, and as shown in fig. 9, the model includes a translation model and a CNN model. As shown in fig. 9, if the triplet vectors of two triples input to the translation model are the same, and the relationship between the head entity and the entity is the same, the tail entity t of the two triples can be determined 1 And t 2 If the degree of difference is large, r is used 1 Representing the entity relationship of the first triple, re-determining the entity relationship in the second triple to obtain r 2 According to r 2 The triplet vectors are retrained. If the difference is small, then according to r 1 And (5) training.
Step 304: and determining second prediction scores of the k first items according to the first entity vectors corresponding to the k first items and the item behavior data of the target user.
Specifically, the similarity between the first entity vectors corresponding to every two of the k first items can be determined, the similarity between every two of the k first items is determined according to the similarity between the first entity vectors corresponding to every two of the k first items, the item similarity matrix of the k first items is determined according to the similarity between every two of the k first items, and finally, the second prediction scores of the k first items are determined according to the item similarity matrix of the k first items and the item behavior data of the target user. The article similarity matrix of the k first articles comprises the similarity of every two articles in the k first articles.
Step 305: recommending the k first items to the target user if the mean square error between the first predicted scores of the k first items and the second predicted scores of the k first items is less than or equal to the error threshold.
It should be noted that, when the mean square error between the first prediction scores of the k first items and the second prediction scores of the k first items is less than or equal to the error threshold, it indicates that when the k first items not only structurally meet the recommendation requirement of the target user, but also semantically meet the recommendation requirement of the target user, at this time, the k first items meeting the similarity between the structural aspect and the semantic aspect may be recommended to the user.
Step 306: and if the mean square error between the first prediction scores of the k first items and the second prediction scores of the k first items is larger than the error threshold, updating the collaborative filtering model according to the second prediction scores of the k first items.
As an example, for the updating of the collaborative filtering model, the β shrinkage factor in the formula (3) may be adjusted to adjust the similarity weight, and then the similarity matrix of multiple articles in the article set is re-determined, so as to achieve the effect of updating the collaborative filtering model. In addition, deeper data mining can be performed on the article behavior data of the user, the calculation process of the similarity in the formula (2) is re-determined according to implicit feedback of more target users, and the effect of updating the collaborative filtering model is achieved. In the embodiment of the present application, no limitation is imposed on the update parameters of the collaborative filtering model.
In addition, if the mean square error between the first prediction scores of the k first items and the second prediction scores of the k first items is greater than the error threshold, a cross-learning task can also be started. The cross learning task may be a task of continuously updating the collaborative filtering model according to the second prediction scores of the k first items, or may be a task of continuously training the triplet vectors through the translation model, which is not limited in the embodiment of the present application. In addition, the cross learning task may also be a task of alternately updating or training the collaborative filtering model and the triple vector, which is not limited in the embodiment of the present application. That is, in the embodiment of the present application, after the cross learning task is started, parameters of the collaborative filtering model may be fixed, model parameters of the translation model may be trained, model parameters of the translation model may also be fixed, and parameters of the collaborative filtering model may be trained, or of course, the two may also be trained alternately until convergence.
Step 307: and determining a third prediction score of the plurality of articles through the updated collaborative filtering model according to the article behavior data of the target user.
In a possible implementation manner, according to the article behavior data of the target user, a similarity matrix of a plurality of articles in the article set is determined through the updated collaborative filtering model, and a third prediction score of the plurality of articles is determined according to the similarity matrix.
Step 308: determining k second items to be recommended from the plurality of items according to the third prediction scores of the plurality of items.
Specifically, first entity vectors corresponding to the k second products respectively can be determined according to the k second products and the knowledge graph vector set; determining fourth prediction scores of the k second articles according to the first entity vectors corresponding to the k second articles respectively and the article behavior data of the target user; recommending the k second items to the target user if the mean square error between the third predicted scores of the k second items and the fourth predicted scores of the k second items is less than or equal to an error threshold; if the mean square error between the third prediction scores of the k second articles and the fourth prediction scores of the k second articles is larger than the error threshold, the collaborative filtering model is continuously updated according to the fourth prediction scores of the k second articles, and the recommended articles are re-determined according to the updated collaborative filtering model until the mean square error between the prediction scores of the articles output by the collaborative filtering model and the predicted scores of the articles output by the collaborative filtering model is smaller than or equal to the error threshold.
In the embodiment of the application, first prediction scores of a plurality of articles are determined through a collaborative filtering model according to article behavior data of a target user, k first articles to be recommended are determined according to the first prediction scores of the plurality of articles, then first entity vectors of the k first articles are determined according to the k first articles and a knowledge graph vector set, second prediction scores of the k first articles are determined according to the first entity vectors of the k first articles and the article behavior data of the target user, and if a mean square error between the first prediction scores and the second prediction scores of the k first articles is smaller than or equal to an error threshold, the k first articles are recommended to the target user. That is, when recommending articles through the collaborative filtering recommendation model, the entity vector of the article in the knowledge graph can be used as auxiliary information for recommendation, so that on the basis of determining the similarity of the articles according to the article behavior data, the similarity between the article attributes is fully considered, the article is accurately predicted and scored by combining the article behavior data of the user and the self-attributes of the article, the recommendation accuracy can be improved, and the problems of cold start and sparseness existing in the collaborative filtering recommendation algorithm are avoided.
Fig. 10 is a schematic structural diagram of an article recommendation apparatus according to an embodiment of the present application, and as shown in fig. 10, the apparatus 1000 includes:
a first determining module 1001, configured to determine, according to item behavior data of a target user, a first prediction score of a plurality of items in an item set through a collaborative filtering model, where the first prediction score of each item is used to indicate a degree of interest of the target user in each item;
a second determining module 1002, configured to determine, according to the first prediction scores of the multiple items, k first items to be recommended to a target user from the multiple items, where k is a positive integer;
a third determining module 1003, configured to determine, according to the k first items and a knowledge graph vector set, first entity vectors corresponding to the k first items, respectively, where the knowledge graph vector set includes first entity vectors of each entity in a knowledge graph, and the knowledge graph includes a plurality of entities that are in one-to-one correspondence with the plurality of items, and entity relationships between every two entities in the plurality of entities;
a fourth determining module 1004, configured to determine second prediction scores of the k first items according to the first entity vectors corresponding to the k first items, respectively, and the item behavior data of the target user;
a first recommending module 1005, configured to recommend the k first items to the target user if a mean square error between the first prediction scores of the k first items and the second prediction scores of the k first items is less than or equal to an error threshold.
Optionally, the fourth determining module includes:
the first determining submodule is used for determining the similarity of every two items in the k first items according to the similarity between the first entity vectors corresponding to every two items in the k first items;
the second determining submodule is used for determining an article similarity matrix of the k first articles according to the similarity of every two articles in the k first articles, wherein the article similarity matrix of the k first articles comprises the similarity of every two articles in the k first articles;
and the third determining submodule is used for determining second prediction scores of the k first items according to the item similarity matrix of the k first items and the item behavior data of the target user.
Optionally, the apparatus further comprises:
the updating module is used for updating the collaborative filtering model according to the second prediction scores of the k first articles if the mean square error between the first prediction scores of the k first articles and the second prediction scores of the k first articles is larger than an error threshold;
the fifth determining module is used for determining third prediction scores of the multiple articles through the updated collaborative filtering model according to the article behavior data of the target user;
a sixth determining module, configured to determine k second items to be recommended from the multiple items according to third prediction scores of the multiple items;
and the second recommending module is used for recommending the object to the target user according to the k second objects.
Optionally, the second recommending module includes:
the fourth determining sub-module is used for determining first entity vectors corresponding to the k second articles according to the k second articles and the knowledge graph vector set;
the fifth determining submodule is used for determining fourth prediction scores of the k second articles according to the first entity vectors corresponding to the k second articles respectively and the article behavior data of the target user;
and the recommending submodule is used for recommending the k second articles to the target user if the mean square error between the third prediction scores of the k second articles and the fourth prediction scores of the k second articles is less than or equal to the error threshold.
Optionally, the knowledge graph includes a plurality of triples, each triplet including a head entity, a tail entity, and an entity relationship between the head entity and the tail entity, the head entity and the tail entity being entities corresponding to any two of the plurality of items;
the device, still include:
a seventh determining module, configured to determine a vector representation of each entity in the multiple triples in the knowledge graph in an entity space, and a vector representation of each entity relationship in the multiple triples in a relationship space, to obtain an initial triplet vector of the multiple triples, where the initial triplet vector of each triplet includes the second entity vector of the head entity, the second entity vector of the tail entity, and the second relationship vector of the entity relationship in each triplet;
the first training module is used for training the initial triple vectors of the multiple triples through the translation model to obtain target triple vectors of the multiple triples, wherein the target triple vector of each triplet comprises a first entity vector of a head entity, a first entity vector of a tail entity and a first relation vector of an entity relation in each triplet, the distance between a reference vector in each triplet and the first entity vector of the tail entity is smaller than a distance threshold, and the reference vector of each triplet refers to the vector sum of the first entity vector of the head entity and the first relation vector of the entity relation in each triplet.
Optionally, the apparatus further comprises:
the optimization module is used for optimizing at least one triplet in the similar triplet set to obtain at least one optimized triplet and determining the initial triplet vector of the at least one optimized triplet if the similar triplet set exists in the triplets in the process of training the initial triplet vectors of the triplets through the translation model; the initial triplet vectors of all triples in the similar triplet set are the same, the head entity and entity relationship in each triplet are correspondingly the same, and the difference degree between tail entities is greater than a difference degree threshold value;
and the second training module is used for training the initial triple vector of the at least one optimized triple through the translation model to obtain a target triple vector of the at least one optimized triple.
Optionally, the optimization module comprises:
a sixth determining sub-module, configured to re-determine, for a reference triple in the at least one triple, an entity relationship between a head entity and a tail entity in the reference triple to obtain an optimized entity relationship, where the reference triple is any triple in the at least one triple;
a seventh determining sub-module, configured to determine, as an optimized triplet corresponding to the reference triplet, the optimized entity relationship and the head entity and the tail entity in the reference triplet;
the feature extraction submodule is used for extracting features of the optimized entity relationship to obtain an optimized relationship vector of the optimized entity relationship;
and the eighth determining submodule is used for determining the optimized relationship vector of the optimized entity relationship as the initial triplet vector of the optimized triplet by referring to the second entity vector of the head entity and the second entity vector of the tail entity in the triplet.
Optionally, the feature extraction sub-module includes:
the obtaining subunit is used for obtaining the description information corresponding to the optimized entity relationship;
the first determining subunit is used for determining the text keywords in the description information;
the second determining subunit is used for determining word vectors of the text keywords through a word vector model according to the text keywords, and the word vector model is used for determining the word vectors of any participle;
the convolution submodule is used for performing convolution processing on the word vectors of the text keywords to obtain a convolution result;
and the pooling submodule is used for pooling the convolution result to obtain an optimized relation vector.
In the embodiment of the application, first prediction scores of a plurality of articles are determined through a collaborative filtering model according to article behavior data of a target user, k first articles to be recommended are determined according to the first prediction scores of the plurality of articles, then first entity vectors of the k first articles are determined according to the k first articles and a knowledge graph vector set, second prediction scores of the k first articles are determined according to the first entity vectors of the k first articles and the article behavior data of the target user, and if a mean square error between the first prediction scores and the second prediction scores of the k first articles is smaller than or equal to an error threshold, the k first articles are recommended to the target user. That is, when recommending articles through the collaborative filtering recommendation model, the entity vector of the article in the knowledge graph can be used as auxiliary information for recommendation, so that on the basis of determining the similarity of the articles according to the article behavior data, the similarity between the article attributes is fully considered, the article is accurately predicted and scored by combining the article behavior data of the user and the self-attributes of the article, the recommendation accuracy can be improved, and the problems of cold start and sparseness existing in the collaborative filtering recommendation algorithm are avoided.
It should be noted that: in the article recommending device provided in the above embodiment, only the division of the functional modules is illustrated when article recommendation is performed, and in practical applications, the function distribution may be completed by different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the article recommendation device and the article recommendation method provided by the above embodiments belong to the same concept, and specific implementation processes thereof are detailed in the method embodiments and are not described herein again.
Fig. 11 is a schematic structural diagram of an electronic device 1100 according to an embodiment of the present application, where the electronic device 1100 may be a terminal or a server, the terminal may be a mobile phone, a tablet computer, or a computer, and the server may be a background server of an item recommendation platform. The electronic device 1100 may have a relatively large difference due to different configurations or performances, and may include one or more processors (CPUs) 1101 and one or more memories 1102, where the memory 1102 stores therein at least one instruction, and the at least one instruction is loaded and executed by the processors 1101 to implement the item recommendation method provided by the above-mentioned method embodiments. Of course, the electronic device 1100 may also have components such as a wired or wireless network interface, a keyboard, and an input/output interface, so as to perform input/output, and the electronic device 1100 may also include other components for implementing device functions, which are not described herein again.
In an exemplary embodiment, a computer-readable storage medium is also provided, which has instructions stored thereon, which when executed by a processor, implement the above item recommendation method.
In an exemplary embodiment, a computer program product is also provided for implementing the above item recommendation method when the computer program product is executed.
It will be understood by those skilled in the art that all or part of the steps for implementing the above embodiments may be implemented by hardware, or may be implemented by a program instructing relevant hardware, where the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.
The above description is only exemplary of the present application and should not be taken as limiting, as any modification, equivalent replacement, or improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims (8)

1. A method for recommending items, the method comprising:
determining first prediction scores of a plurality of items in an item set through a collaborative filtering model according to item behavior data of a target user, wherein the first prediction score of each item is used for indicating the interest degree of the target user in each item;
determining k first items to be recommended to the target user from the plurality of items according to the first prediction scores of the plurality of items, wherein k is a positive integer;
determining first entity vectors corresponding to the k first articles respectively according to the k first articles and a knowledge graph vector set, wherein the knowledge graph vector set comprises first entity vectors of all entities in a knowledge graph, and the knowledge graph comprises a plurality of entities corresponding to the articles one by one and entity relations between every two entities in the entities;
determining second prediction scores of the k first items according to first entity vectors corresponding to the k first items respectively and item behavior data of the target user;
recommending the k first items to the target user if the mean square error between the first prediction scores of the k first items and the second prediction scores of the k first items is less than or equal to an error threshold;
if the mean square error between the first prediction scores of the k first items and the second prediction scores of the k first items is larger than the error threshold value, starting a cross learning task, wherein the cross learning task comprises updating the collaborative filtering model according to the second prediction scores of the k first items;
the knowledge graph comprises a plurality of triples, each triplet comprises a head entity, a tail entity and an entity relationship between the head entity and the tail entity, and the head entity and the tail entity are entities corresponding to any two articles in the plurality of articles;
before determining the first entity vectors corresponding to the k first items respectively according to the k first items and the knowledge graph vector set, the method further includes:
determining vector representations of all entities in a plurality of triples in an entity space and vector representations of all entity relationships in the triples in a relationship space, which are included in the knowledge graph, to obtain initial triplet vectors of the triples, wherein the initial triplet vector of each triplet includes a second entity vector of a head entity, a second entity vector of a tail entity and a second relationship vector of an entity relationship in each triplet; training the initial triplet vectors of the multiple triplets through a translation model to obtain target triplet vectors of the multiple triplets, wherein the target triplet vector of each triplet comprises a first entity vector of a head entity, a first entity vector of a tail entity and a first relation vector of an entity relation in each triplet, the distance between a reference vector in each triplet and the first entity vector of the tail entity is smaller than a distance threshold, and the reference vector is the vector sum of the first entity vector of the head entity and the first relation vector of the entity relation;
the method further comprises the following steps:
in the process of training the initial triplet vectors of the multiple triples through the translation model, if a similar triplet set exists in the multiple triples, re-determining an entity relationship between a head entity and a tail entity in a reference triplet of at least one triplet in the similar triplet set to obtain an optimized entity relationship, where the reference triplet is any triplet in the at least one triplet; determining the optimized entity relationship and a head entity and a tail entity in the reference triple as an optimized triple corresponding to the reference triple; performing feature extraction on the optimized entity relationship to obtain an optimized relationship vector of the optimized entity relationship; determining an optimization relationship vector of the optimization entity relationship, a second entity vector of a head entity and a second entity vector of a tail entity in the reference triplet as an initial triplet vector of the optimization triplet; the initial triplet vectors of all triples in the similar triplet set are the same, the head entity and entity relationship in each triplet are correspondingly the same, and the difference degree between tail entities is greater than a difference degree threshold value; and training the initial triple vector of at least one optimized triple through the translation model to obtain the target triple vector of the at least one optimized triple.
2. The method according to claim 1, wherein the determining second prediction scores for the k first items according to the first entity vectors corresponding to the k first items, respectively, and the item behavior data of the target user comprises:
determining the similarity of every two items in the k first items according to the similarity between the first entity vectors corresponding to every two items in the k first items;
determining an article similarity matrix of the k first articles according to the similarity of every two articles in the k first articles, wherein the article similarity matrix of the k first articles comprises the similarity of every two articles in the k first articles;
and determining second prediction scores of the k first items according to the item similarity matrix of the k first items and the item behavior data of the target user.
3. The method of claim 1, further comprising:
determining third prediction scores of the multiple articles through the updated collaborative filtering model according to the article behavior data of the target user;
determining k second items to be recommended from the plurality of items according to the third prediction scores of the plurality of items;
and recommending the object to the target user according to the k second objects.
4. The method of claim 3, wherein said recommending items to the target user based on the k second items comprises:
determining first entity vectors corresponding to the k second products respectively according to the k second products and the knowledge graph vector set;
determining fourth prediction scores of the k second articles according to the first entity vectors corresponding to the k second articles respectively and the article behavior data of the target user;
recommending the k second items to the target user if the mean square error between the third prediction scores of the k second items and the fourth prediction scores of the k second items is less than or equal to the error threshold.
5. The method of claim 1, wherein the performing feature extraction on the optimized entity relationship to obtain an optimized relationship vector of the optimized entity relationship comprises:
obtaining description information corresponding to the optimized entity relationship;
determining text keywords in the description information;
determining word vectors of the text keywords through a word vector model according to the text keywords, wherein the word vector model is used for determining the word vectors of any participle;
performing convolution processing on the word vectors of the text keywords to obtain a convolution result;
and performing pooling treatment on the convolution result to obtain the optimization relation vector.
6. An item recommendation device, the device comprising:
the first determining module is used for determining first prediction scores of a plurality of articles in an article set through a collaborative filtering model according to article behavior data of a target user, wherein the first prediction score of each article is used for indicating the interest degree of the target user in each article;
a second determining module, configured to determine, according to the first prediction scores of the multiple items, k first items to be recommended to the target user from the multiple items, where k is a positive integer;
a third determining module, configured to determine first entity vectors corresponding to the k first items according to the k first items and a knowledge graph vector set, where the knowledge graph vector set includes first entity vectors of entities in a knowledge graph, and the knowledge graph includes a plurality of entities that correspond to the items one to one and entity relationships between every two entities in the entities;
a fourth determining module, configured to determine second prediction scores of the k first items according to the first entity vectors corresponding to the k first items, respectively, and the item behavior data of the target user;
a first recommending module, configured to recommend the k first items to the target user if a mean square error between the first prediction scores of the k first items and the second prediction scores of the k first items is less than or equal to an error threshold; if the mean square error between the first prediction scores of the k first items and the second prediction scores of the k first items is larger than the error threshold, starting a cross learning task, wherein the cross learning task comprises updating the collaborative filtering model according to the second prediction scores of the k first items;
the knowledge graph comprises a plurality of triples, each triplet comprises a head entity, a tail entity and an entity relationship between the head entity and the tail entity, and the head entity and the tail entity are entities corresponding to any two articles in the plurality of articles;
the apparatus also includes means for:
determining vector representations of all entities in a plurality of triples in an entity space and vector representations of all entity relationships in the triples in a relationship space, which are included in the knowledge graph, to obtain initial triplet vectors of the triples, wherein the initial triplet vector of each triplet includes a second entity vector of a head entity, a second entity vector of a tail entity and a second relationship vector of an entity relationship in each triplet; training the initial triplet vectors of the multiple triplets through a translation model to obtain target triplet vectors of the multiple triplets, wherein the target triplet vector of each triplet comprises a first entity vector of a head entity, a first entity vector of a tail entity and a first relation vector of an entity relation in each triplet, the distance between a reference vector in each triplet and the first entity vector of the tail entity is smaller than a distance threshold, and the reference vector is the vector sum of the first entity vector of the head entity and the first relation vector of the entity relation;
the apparatus also includes means for:
in the process of training the initial triplet vectors of the multiple triples through the translation model, if a similar triplet set exists in the multiple triples, re-determining an entity relationship between a head entity and a tail entity in a reference triplet of at least one triplet in the similar triplet set to obtain an optimized entity relationship, where the reference triplet is any triplet in the at least one triplet; determining the optimized entity relationship and a head entity and a tail entity in the reference triple as an optimized triple corresponding to the reference triple; performing feature extraction on the optimized entity relationship to obtain an optimized relationship vector of the optimized entity relationship; determining an optimization relationship vector of the optimization entity relationship, a second entity vector of a head entity and a second entity vector of a tail entity in the reference triplet as an initial triplet vector of the optimization triplet; the initial triplet vectors of all triples in the similar triplet set are the same, the head entity and entity relationship in each triplet are correspondingly the same, and the difference degree between tail entities is greater than a difference degree threshold value; and training the initial triple vector of at least one optimized triple through the translation model to obtain the target triple vector of the at least one optimized triple.
7. An electronic device, characterized in that the electronic device comprises:
a processor;
a memory for storing processor-executable instructions;
wherein the processor is configured to perform the steps of the method of any of the above claims 1 to 5.
8. A computer readable storage medium having instructions stored thereon, wherein the instructions, when executed by a processor, implement the steps of any one of the method of claims 1 to 5.
CN202010224857.7A 2020-03-26 2020-03-26 Article recommendation method and device and computer storage medium Active CN113449200B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010224857.7A CN113449200B (en) 2020-03-26 2020-03-26 Article recommendation method and device and computer storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010224857.7A CN113449200B (en) 2020-03-26 2020-03-26 Article recommendation method and device and computer storage medium

Publications (2)

Publication Number Publication Date
CN113449200A CN113449200A (en) 2021-09-28
CN113449200B true CN113449200B (en) 2023-04-18

Family

ID=77807250

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010224857.7A Active CN113449200B (en) 2020-03-26 2020-03-26 Article recommendation method and device and computer storage medium

Country Status (1)

Country Link
CN (1) CN113449200B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116089624B (en) * 2022-11-17 2024-02-27 昆仑数智科技有限责任公司 Knowledge graph-based data recommendation method, device and system
CN115828771B (en) * 2023-02-13 2023-04-28 深圳市仕瑞达自动化设备有限公司 Performance evaluation method, system and medium for mechanical transmission element

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107729444A (en) * 2017-09-30 2018-02-23 桂林电子科技大学 Recommend method in a kind of personalized tourist attractions of knowledge based collection of illustrative plates
CN108733798A (en) * 2018-05-17 2018-11-02 电子科技大学 A kind of personalized recommendation method of knowledge based collection of illustrative plates
CN109191240A (en) * 2018-08-14 2019-01-11 北京九狐时代智能科技有限公司 A kind of method and apparatus carrying out commercial product recommending
CN109783738A (en) * 2019-01-22 2019-05-21 东华大学 A kind of double extreme learning machine mixing collaborative filtering recommending methods based on more similarities

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107729444A (en) * 2017-09-30 2018-02-23 桂林电子科技大学 Recommend method in a kind of personalized tourist attractions of knowledge based collection of illustrative plates
CN108733798A (en) * 2018-05-17 2018-11-02 电子科技大学 A kind of personalized recommendation method of knowledge based collection of illustrative plates
CN109191240A (en) * 2018-08-14 2019-01-11 北京九狐时代智能科技有限公司 A kind of method and apparatus carrying out commercial product recommending
CN109783738A (en) * 2019-01-22 2019-05-21 东华大学 A kind of double extreme learning machine mixing collaborative filtering recommending methods based on more similarities

Also Published As

Publication number Publication date
CN113449200A (en) 2021-09-28

Similar Documents

Publication Publication Date Title
WO2020211566A1 (en) Method and device for making recommendation to user, computing apparatus, and storage medium
CN110162706B (en) Personalized recommendation method and system based on interactive data clustering
CN108920641B (en) Information fusion personalized recommendation method
CN109033101B (en) Label recommendation method and device
CN111553754A (en) Updating method and device of behavior prediction system
CN110197404B (en) Personalized long-tail commodity recommendation method and system capable of reducing popularity deviation
CN112417306A (en) Method for optimizing performance of recommendation algorithm based on knowledge graph
CN108763367B (en) Method for recommending academic papers based on deep alignment matrix decomposition model
CN113449200B (en) Article recommendation method and device and computer storage medium
CN111949887A (en) Item recommendation method and device and computer-readable storage medium
CN112256965A (en) Neural collaborative filtering model recommendation method based on lambdamat
CN112380451A (en) Favorite content recommendation method based on big data
CN112085158A (en) Book recommendation method based on stack noise reduction self-encoder
CN115618101A (en) Streaming media content recommendation method and device based on negative feedback and electronic equipment
CN110085292A (en) Drug recommended method, device and computer readable storage medium
CN111259223B (en) News recommendation and text classification method based on emotion analysis model
CN111198991A (en) Collaborative filtering recommendation method based on trust level and expert user
Pawar et al. Movies Recommendation System using Cosine Similarity
CN116186413A (en) Method and system for multi-modal feature recommendation
CN112650869B (en) Image retrieval reordering method and device, electronic equipment and storage medium
CN114936901A (en) Visual perception recommendation method and system based on cross-modal semantic reasoning and fusion
CN113297485A (en) Method for generating cross-modal representation vector and cross-modal recommendation method
CN114238752A (en) Article recommendation method and device and storage medium
CN113763084A (en) Product recommendation processing method, device, equipment and storage medium
KR20210120977A (en) Interactive customized search method based on limited Boltzmann machine drive

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant