CN112733027B - Hybrid recommendation method based on local and global representation model joint learning - Google Patents
Hybrid recommendation method based on local and global representation model joint learning Download PDFInfo
- Publication number
- CN112733027B CN112733027B CN202110026008.5A CN202110026008A CN112733027B CN 112733027 B CN112733027 B CN 112733027B CN 202110026008 A CN202110026008 A CN 202110026008A CN 112733027 B CN112733027 B CN 112733027B
- Authority
- CN
- China
- Prior art keywords
- information
- local
- user
- global
- graph
- 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
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
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- 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
- 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
- G06F16/958—Organisation or management of web site content, e.g. publishing, maintaining pages or automatic linking
-
- 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
-
- 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
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)
- Life Sciences & Earth Sciences (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Animal Behavior & Ethology (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a hybrid recommendation algorithm based on local and global representation model joint learning, which comprises the following steps: adopting entity link to link the project with an entity in an external knowledge graph and obtain text description content and heterogeneous relation in the project as local information and global information respectively; constructing a knowledge complement embedded network as a local model; constructing a knowledge perception graph convolutional neural network as a global model; and integrating the representation information of the user and the project in the local embedding space and the global embedding space respectively, inputting the integrated result as the characteristic into the multilayer perceptron for prediction, outputting the required result, and finishing the joint learning process. The algorithm can capture the preference information of the user to the project at different angles, further effectively predict the click probability of the user, has high accuracy of the prediction result and small error, can be applied to various recommendation scenes, and has high practical value.
Description
Technical Field
The invention belongs to the field of recommendation systems, and particularly relates to a hybrid recommendation method based on local and global representation model joint learning.
Background
With the development of internet technology, the total amount of data of online information rises sharply, and excessive information makes users unable to obtain the information really useful for themselves when facing a large amount of information, i.e. information overload.
Traditional collaborative filtering methods typically employ user-project-based historical interaction data to solve the problem, but these methods typically face data sparsity and cold start issues. In recent years, many methods solve the problems by combining social network information of users and attribute information of items as auxiliary information, but most of the methods face the problems of low data quality, poor modeling method effect, failure to effectively mine effective information in the social network information and the item information, and the like.
With the rapid development of semantic networks, knowledge graph data containing rich heterogeneous relationships provides an opportunity to know the relevant relations among users, projects and user-projects, and rules and mutual influences among nodes in a graph can be revealed. How to effectively model the preference information of the user through comprehensive research on utilizing rich heterogeneous data in the knowledge graph is a research focus for realizing personalized recommendation.
Disclosure of Invention
The invention aims to model user preference information in a recommendation environment and further predict the problem of the click probability of a user on an item, and provides a mixed recommendation method based on local and global representation model joint learning, which can comprehensively consider local information and global information related to the item and aggregate historical interaction information of the user to model the preference information of the user.
In order to achieve the purpose, the invention is realized by the following technical scheme:
a hybrid recommendation method based on joint learning of local and global representation models comprises the following steps:
step 1: by adopting entity link, the project is linked with an entity in an external knowledge graph, and text description content and heterogeneous relation in the project are acquired and respectively used as local information and global information;
and 2, step: constructing a knowledge complement embedded network as a local model according to the local information;
and step 3: constructing a knowledge perception graph convolutional neural network as a global model according to global information;
and 4, step 4: and (3) considering the historical interactive behavior of the user, simulating the local preference and the global preference of the user to the item under different models by combining an attention mechanism, integrating the representation information of the user and the item in local and global embedding spaces respectively, inputting the integrated result as a feature into a multi-layer perception machine for prediction, outputting a required result, and finishing the joint learning process.
Preferably, in step 1, the names of the items are used as queries and the type information of the items is used as constraint conditions, the items are linked with the entities in the external knowledge graph, and a heterogeneous relationship information graph between the text description information related to the items and the entities is obtained.
Preferably, obtaining a heterogeneous relationship information graph between the text description information related to the project and the entity includes:
step 1.1: the method comprises the steps that interactive information between a user and a project and basic information of the project are obtained in a centralized mode in project recommendation data;
step 1.2: respectively representing users in the data set, items in the data set, the number of the items and type information of the items by using a list;
step 1.3: taking the title of the project as query content, searching API by using an offline knowledge graph, and retrieving KB entity from the knowledge graph Freebase;
step 1.4: based on the linked entity, text description information is extracted from the KB, description information corresponding to the item is represented by a list, structure information is extracted, and representation is represented by a heterogeneous graph.
Preferably, the data comprises historical interaction data between the user and the item; the project information comprises a title and a category; the local information and the global information respectively comprise entity description information and heterogeneous relation information in the knowledge graph.
Preferably, in the step 2, a knowledge enhanced embedding algorithm is used for the text description content of the item to obtain a local embedded representation of the item; and calculating the weight value to represent the correlation degree between the two projects, constructing a historical interactive binary graph structure according to the interactive behavior information between the user and the projects, and calculating by combining the two to obtain the local embedded representation of the user.
Preferably, the obtaining of the locally embedded representation of the item using a knowledge enhanced embedding algorithm on the textual description of the item comprises:
step 2.1: aligning the title of the item with the text description information obtained in the step 1.4; obtaining each word and the initial embedded representation thereof;
step 2.2: calculating a title-embedded representation of the obtained title and text description using a knowledge enhanced embedding algorithmAnd textual description sentence embedding representationFinally, the embedded representations of the items are merged to obtain the local embedded representation of the items.
Preferably, a history interaction binary graph structure is constructed, and the user local embedded representation is obtained by combining the history interaction binary graph structure and the user local embedded representation through calculation, wherein the method comprises the following steps:
step 2.3: representing user u using lists based on locally embedded representations of items i Setting a specific weight calculation mode, and respectively calculating weights between the historical interactive items and the candidate items;
step 2.4: and corresponding the weight to the historical interactive item to represent the contribution degree to modeling user preference information, and combining the weight with the local embedded representation to obtain a new embedded representation which indicates the local embedded representation of the user.
Preferably, the step 3, obtaining a global embedded representation of the item and the user by using a knowledge-graph convolutional neural network for the structural information in the knowledge-graph, includes:
step 3.1: combining the user, the project and the interaction information Y between the user and the project to add the user, the project and the interaction information Y into the abnormal graph H obtained in the step 1.4 to obtain a unified relational graph G;
step 3.2: using an embedded learning algorithm based on transfer to the relationship graph G constructed in the step 3.1 to obtain embedded representation of nodes and relationships in the relationship graph G;
step 3.3: taking a user node in the relational graph G as a central node, and taking a directly connected node as a neighbor node; aggregating the neighbor node characteristics in the relational graph G by adopting a knowledge-aware graph convolution neural network to obtain neighbor embedding representation and obtain new embedding by combining with the embedding information of the user as the user u i Is embedded in the information.
Preferably, step 4 specifically comprises the following steps:
step 4.1: user u i And candidate item v i The local embedding expresses that the cosine similarity is calculated to obtain the local click probability p l ;
Step 4.2, user u i And candidate item v i The local embedding expresses that the cosine similarity is calculated to obtain the local click probability p g ;
Step 4.3: the local click probability p calculated in the steps 4.1 and 4.2 l And global click probability p g And aggregating through the weight calculated by the door mechanism to obtain the user u i For candidate item v i Establishing a corresponding target function, and training by using an optimizer to minimize a loss function;
step 4.4: and after one training is finished, updating the behavior and site node potential characteristics, and circulating the calculation processes in the steps 3, 4 and 5 to carry out the next training and result output.
The method solves the problem of cold start of the project in the recommendation environment by introducing the related auxiliary information of the project; according to the degree of closeness of contact with the project, the information related to the project is innovatively divided into local information and global information; in the aspect of information acquisition, the project is linked with the entity in the external knowledge graph by adopting an entity linking technology, and the text description content and the heterogeneous relation in the project are acquired as local information and global information respectively. According to the characteristics of different types of information, a knowledge complement embedded network is respectively designed as a local model and a knowledge perception graph convolutional neural network is respectively designed as a global model for mining different characteristics in the local model and the global model. In the joint learning process, the historical interaction behaviors of the user are considered firstly, and the local preference and the global preference of the user on the items are simulated under different models by combining an attention mechanism. And finally, aggregating different preference information as characteristics to be input into the multilayer perceptron for prediction, aggregating different information through a door mechanism, and calculating to obtain the click probability of the user to the item. And effective modeling of the actual preference information of the user is completed, and the personalized recommendation process of the user is realized.
Due to the adoption of the technical scheme, the invention has the following beneficial effects:
(1) According to the method, the information is classified, the project related information is divided into local information and global information based on the interactive object, and the preference information of the user can be mined from different angles by modeling different information;
(2) According to the method, projects are linked with entities in an external knowledge graph through an entity linking technology, then rich heterogeneous relationships among the entities are extracted, and further modeling is performed on preference information of a user from multiple angles by combining description information of the projects and historical interaction information between the user and the projects;
(3) According to the method, the local and global information of the user/project is mined by designing different deep learning architectures to obtain the local preference and the global preference of the user to the project, and the local preference and the global preference are aggregated to be input into the multi-layer perceptron as features to carry out co-learning and prediction of the user click probability, so that the prediction result has higher accuracy, smaller error and higher practical value.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the invention and together with the description serve to explain the principles of the invention:
FIG. 1 is a flow diagram of a hybrid recommendation algorithm based on joint learning of local and global representation models;
FIG. 2 is a diagram illustrating the acquisition of local information and global information related to a project by an entity linking method;
FIG. 3 is a knowledge enhancement embedding method in conjunction with local information;
FIG. 4 is a method of aggregating historical interactive movie computations to obtain a user local embedded representation;
FIGS. 5 (a), 5 (b) show a unified relationship diagram and a knowledge-aware convolutional neural network, respectively;
fig. 6 (a) and 6 (b) are respectively effectiveness visualizations of the hybrid recommendation algorithm based on the local and global representation model joint learning under different parameters.
Detailed Description
The present invention will now be described in detail with reference to the drawings and specific embodiments, wherein the exemplary embodiments and descriptions of the present invention are provided to explain the present invention without limiting the invention thereto.
As shown in FIG. 1, the hybrid recommendation method based on the joint learning of the local and global representation models of the present invention includes the following steps:
step 1: by adopting entity link, the project is linked with the entity in the external knowledge map, and the text description content and the heterogeneous relation in the project are acquired as local information and global information respectively.
Through an entity linking process, the name of a project is used as query and information such as the type of the project is used as constraint conditions, the project is linked with entities in an external knowledge graph (KB), and text description information related to the project is obtained and used as local information and a heterogeneous relation information graph between the entities and used as global information.
The method specifically comprises the following steps:
step 1.1: and intensively acquiring interaction information between the user and the item and basic information of the item in the item recommendation data.
Step 1.2: using list U = { U = 1 ,u 2 ,…,u M Denotes a user in the data set, where u i E, representing the ith user in the list U by the U, wherein M is the number of the users; list V = { V = 1 ,v 2 ,…,v N Denotes an item in the data set, where v j The element V represents the jth commodity in the list V, and N is the quantity of the commodities; listsA title representing item i, where n represents the title length; character string t i Indicating the type information of the article i. Matrix Y ∈ R whose elements are 0,1 M×N Representing interaction data between a user and a good, where y uv =1 indicates that there is an interaction between user u and item i, otherwise y uv =0;
Step 1.3: title W of commodity i i As query content, KB entities are retrieved from the knowledge-graph Freebase using the offline knowledge-graph search API.
For items that return more than one search entity as a result during the entity linking process, an item title W is further employed i And type T i Screening as constraint conditions to obtain accurate link entity e i 。
Step 1.4: for linking entity e i Extracting text description information of the knowledge graph as a project v i Using lists of local informationWherein m represents a description information length; extracting the heterogeneous relation information thereof as a project v i Using an abnormal picture Representation of a triplet (e) h ,e r ,e t ) Indicating the presence of the de novo node e h To tail node e t Relation e of r ξ andrepresenting a list of nodes and relationships, respectively.
The data in step 1.1 to step 1.4 comprises historical interaction data between the user and the project; the item information includes title, category, etc.; the local information and the global information respectively comprise entity description information, heterogeneous relation information in a knowledge graph and the like.
Step 2: and constructing a knowledge complementing embedded network as a local model according to the local information, and obtaining a local embedded representation of the item by using a knowledge enhancing embedding method for the text description content of the item. The method specifically comprises the following steps;
step 2.1: will item v i Title W of i And the text description information D obtained in step 1.4 i Align, get each word therein using word2vecAndinitial embedded representation ofAnd
step 2.2: calculating a title-embedded representation of the obtained title and text description using a knowledge-enhanced embedding algorithmAnd textual description sentence embedding representationFinally, the embedded representations of the items are merged to obtain an item v i Is locally embedded to representWherein for candidate item v j Its local embedding is expressed as
In step 2.2, the knowledge enhancement embedding algorithm is specifically realized as follows:
knowledge-enhanced embedding algorithm uses a convolutional neural network model to process a sentence s = { w = { a word sequence constitutes 1 ,w 2 …, using filters to construct the word embedded representation, then using max pooling to extract the maximum value in each dimension of the representation, resulting in new representation information as the embedded representation of the sentenceThe process is defined as follows:
wherein W 1 And h represent weight and size information of a filter in the convolutional neural network, respectively.
Title embedded representationAnd textual description sentence embedding representationThe specific implementation of the post-merger computation local embedding representation of (1) is:
the merging process is represented by embedding the title into the representationAnd textual description sentence embedding representationStitching as a deep neural networkIs calculated to obtain a knowledge-enhanced embedded representationAs item v i Is locally embedded to representThe process is defined as follows:
And calculating the weight value to represent the correlation degree between the two projects, constructing a historical interactive binary graph structure according to the interactive behavior information between the user and the projects, and calculating by combining the two to obtain the local embedded representation of the user.
Step 2.3: using the list R (u) based on the locally embedded representation of the item i )={v 1 ,v 2 ,…,v n Denotes user u i And v j Representing candidate items, wherein n is the number of historical interaction items. And setting a specific weight calculation mode, and calculating the weights between the historical interactive items and the candidate items respectively.
The weight calculation between the historical interactive item and the candidate item is specifically realized as follows:
the weight calculation method is that two layers of deep neural networks are usedThe attention mechanism is formed by embedding historical interaction items into a representationAnd candidate item embeddingAs input, the weight values between them are calculatedThen, the weight value is normalized by using a softmax function to obtain
Where R (u) represents the historical set of interactive merchandise for user u, W 1 Andb 1 and b 2 Respectively representWeight and bias in (1).
Step 2.4: corresponding the weight value to the historical interactive item to represent the contribution degree to the modeling user preference information, and combining the weight value with the local embedded representation to aggregate to obtain a new embedded representationIndicating user u i Is shown partially embedded.
The aggregation process of the historical items is specifically realized as follows:
the weight a obtained by combining the calculation of the step 3.1 in the polymerization process kj And corresponding item-embedded representationsComputing a locally embedded representation of a user
Where n represents the historical interactive commodity quantity for user u.
And step 3: a knowledge graph convolutional neural network is used for obtaining global embedded representation of the items and the users for the structural information in the knowledge graph. The method specifically comprises the following steps;
step 3.1: and (4) combining the user, the item and the interaction information Y between the user and the item, and adding the user, the item and the interaction information Y into the abnormal graph H obtained in the step 1.4 to obtain a unified relation graph G.
The process of constructing the unified relationship graph G is specifically realized as follows:
the construction process adds the users, commodities and the mutual information among the users and the commodities into the heterogeneous knowledge graphObtaining a unified relationship diagramWhere ξ' = { ξ utou } indicates that the user is added to the graph as a new node,showing the interaction between the user and the goods as a new relationship to the graph.
Step 3.2: using embedded learning algorithm based on translation to obtain embedded representation of node and relation in graph data G constructed in step 3.1Andfor item v i By usingRepresenting its global embedded information. For user node u i The panel is embeddedThe in representation is its initial embedded representation
The embedded learning algorithm based on the transfer is specifically realized as follows:
the translation-based embedded learning algorithm performs embedded representation learning by optimizing translation rules:
whereinAndrespectively representing the embedded representation of nodes h, t and the relation r in the graph, M r A mapping matrix is represented that is,to representA mapped representation in the space of the relation r. For a triplet (h, r, t), its confidence score is defined as:
g r a lower score of (h, t) indicates that (h, r, t) is more likely to be a valid triplet. Then, the boundary-based ordering loss function is used for training, and the representation information of the nodes and the relations in the graph is obtained through learning:
step 3.3: with user node u in graph G i As a central node, taking a directly connected node as its neighbor node set N (u) i ). Neighbor node characteristics in the graph are aggregated by adopting a knowledge-aware graph convolution neural network, and a user u is combined i Initial embedded representationCalculating to obtain new embedding as its global embedding informationAnd (4) showing.
The knowledge-aware graph convolution neural network and the process of acquiring the global embedded representation are specifically realized as follows:
knowledge-aware graph-convolutional neural networks are used to aggregate embedded representations in a non-linear fashion, using localized convolution operations to merge ego-the embedding of neighbor nodes in the network that are directly connected to the target node h, as defined below:
where σ denotes the nonlinear activation function, N (h) denotes a set π consisting of neighboring nodes directly connected to node h ht The normalized aggregation coefficient between the node h and the node t is represented, different importance of different neighbor nodes to the target node is reflected, and the normalized aggregation coefficient is defined as follows:
simultaneous user node u i By incorporating neighbor embeddingAnd the initial embedded representation of itselfCalculated, defined as follows:
wherein W 6 And LeakyRelu represent weights and activation functions in the neural network, respectively.
And 4, step 4: in the joint learning process, the historical interactive behaviors of the user are considered, the local preference and the global preference of the user to the item are simulated under different models by combining an attention mechanism, the representation information of the user and the item in local and global embedding spaces is integrated respectively, the integrated result is used as a feature and input into a multi-layer perception machine for prediction, and a required result is output. The method specifically comprises the following steps:
step 4.1: user u i And candidate item v j The local embedding expresses that the cosine similarity is calculated to obtain the local click probability p l ;
Step 4.2: user u i And candidate item v j The global embedding expresses that the cosine similarity is calculated to obtain the local click probability p g ;
Step 4.3: for the local click probabilities p calculated in steps 4.1 and 4.2 l And global click probability p g Designing a door mechanism to calculate to obtain a weight coefficient theta between the door mechanism and the door mechanism, and finally performing polymerization calculation to obtain a user u i For candidate item v j The click probability Pr of. Establishing a corresponding objective function, and training by using a proper optimizer to minimize a loss function;
the design of the prediction process and the loss function of the click probability is specifically realized as follows:
local click probability in combination with user u i And candidate item v j Is calculated as follows:
global click probability in conjunction with user u i And candidate item v j Is calculated as follows:
the predicted click probability Pr is calculated through a door mechanism to obtain a weight coefficient lambda between the predicted click probability Pr and the predicted click probability is finally calculated in a polymerization mode, and the predicted click probability Pr is defined as follows:
Pr(u i ,v j )=λp l (u i ,v j )+(1-λ)p g (u i ,v j )
where σ is sigmoid function, W gated A weight matrix of a gate mechanism, representing a vector dot product.
The training process utilizes a cross entropy function as a loss function and adopts mini-batch Adam for optimization, and the definition is as follows:
wherein R is ij For elements in the interaction matrix R, user u is represented i And candidate goods v j Whether there is an interaction between them.
Step 4.4: and after one training is finished, updating the behavior and site node potential characteristics, and circulating the calculation processes in the steps 3, 4 and 5 to carry out the next training and result output.
The invention is further illustrated by the following specific examples.
In the embodiment, the recommendation system data set is MovieLens-20M, and a source of the movie evaluation related data sethttps://grouplens.org/datasets/movielensAfter data preprocessing, 9861984 pieces of evaluation data are obtained, which comprise 61859 different users and 17488 different movies, and 5-star evaluation is adopted. Setting the rating threshold to 3, and setting the rating to be greater than or equal to 3 is regarded as that the preference of the user for the movie is positive, and negative. The used knowledge graph data is the Freebase data of the latest published version, and the triples are used<The head, relative, tail forms store the relevant facts. Referring to fig. 2, a movie item "a movie" in a data set is linked with an entity in an external knowledge graph through an entity linking technology, so as to obtain a related text description and heterogeneous relationship graph information thereof, which are respectively used as local and global information thereof, and the process is as follows:
step 1.2: use list U = { U = { (U) } 1 ,u 2 ,…,u M Denotes a user in the data set, where u i E, representing the ith user in the list U by the U, wherein M is the number of the users; list V = { V = 1 ,v 2 ,…,v N Denotes an item in the dataset, where v i And e.g. V represents the ith item in the list V, and N is the number of items. Taking FIG. 2 as an example, item v i For a movie ", the titles are mapped and presented as a list Wherein n represents the title length with fixed size, fixed as 10 in this example, and filled with less than 0, wherein each element represents the index value corresponding to each word after the word segmentation of the movie title; character string t i Representing a movie v i The type information of (1) is love photos. Matrix Y ∈ R whose elements are 0,1 M×N Representing interaction data between a user and a good, where y ij =1 denotes user u i And commodity v j There is an interactive action between them, otherwise y ij =0;
Step 1.3: using the movie title as a query, and using the entity linking technology to link the movie v i Linking with entities in the knowledge-graph to obtain an item-entity pair < a movie,/m/02 v7x2 >, wherein/m/02 v7x2 represents the id of the entity in the knowledge-graph. Then, the local text description information "a certain movie" is … in 2007 "as shown in fig. 2 is acquired, and the local information is mapped and represented as a list:
wherein m represents the description information length, fixed as 20, and filled with less than 0; and extracting the triple information of the entity [ < http:// rdf. Freebase. Com/ns/m.02v7x2 >, < http:// rdf. Freebase. Com/ns/film.film.cinematography >, < http:// rdf. Freebase. Com/ns/m.0jsqw2 > ]]Form a movie v i And the heterogeneous relationship graph which is the central node is used as global information.
And 2, step: as shown in fig. 3, a knowledge-enhanced embedding algorithm is used on the movie text description content to obtain a locally embedded representation of the item.
Step 2.1: will movie v i Title W of i And the local information D obtained in step 1 i Align, get each word w therein using word2vec ij ∈W i And d ij ∈D i Initial embedded representation of Anda dimension size of 50;
step 2.2: calculating a title-embedded representation of the obtained title and text description using a knowledge-enhanced embedding algorithmAnd textual description sentence embedding representationFinally, the embedded representations are combined to obtain the local embedded representation of the film The dimension size is 50.
According to the graph shown in FIG. 4, a historical interactive binary graph structure is constructed according to the interactive behavior information between the user and the movie, an attention network is designed, weight information is calculated according to the correlation degree of the historical interactive movie and the candidate movie, and finally the historical interactive movie is subjected to aggregation calculation to obtain the local embedded representation of the user;
step 2.3: using the list R (u) based on the locally embedded representation of the movie i )={v 1 ,v 2 ,…,v n Denotes user u i And v j Representing candidate items, wherein n is the fixed number of history interactive items 10. Calculating the weight between the historical interactive item and the candidate itemThe dimension size is 10.
Step 2.4: corresponding the weight value to the historical interactive item to represent the contribution degree to the modeling user preference information, and combining the weight value and the historical interactive item to represent and aggregate the contribution degree to the modeling user preference information to obtain a user u i Is locally embedded to represent The dimension size is 50.
And 3, step 3: a knowledge graph convolutional neural network is used for obtaining global embedded representation of the items and the users for the structural information in the knowledge graph. The method specifically comprises the following steps;
step 3.1: as shown in FIG. 5 (a), the user and the item, and the interaction information Y between them are combined and added to the abnormal image H obtained in step 1.4 to obtain the unified relationship diagram Where ξ' = { ξ ≡ U } indicates that the user is added to the graph as a new node,showing the interaction between the user and the goods as a new relationship to the graph.
Step 3.2: using an embedded learning algorithm based on translation for the graph data G constructed in the step 4.1 to obtain embedded representation of nodes and relations in the graph data GWhere t ∈ ξ' andfor movie v i , Indicating its global embedded information.
Step 3.3: as shown in FIG. 5 (b), with user node u in graph G i And taking the directly connected node as a neighbor node of the node as a central node. Aggregating neighbor node characteristics in the graph G by adopting a knowledge-aware graph convolution neural network to obtain neighbor embedded representationEmbedding information in connection with a userGet a new embedding as user u i Global embedded information of
And 4, step 4: in the process of joint learning, the representation information of the user and the item in the local embedding space and the global embedding space is integrated respectively, the integrated result is used as the characteristic to be input into the multilayer perceptron to be predicted, and the required result is output. The method specifically comprises the following steps:
step 4.1: user u i And candidate item v i Is obtained by splicing The dimension size is 100. Inputting the local click probability p as a characteristic into a multilayer perceptron l (u i ,v j )=0.6157。
Step 4.2 user u i And candidate item v i Is used for splicing to obtain The dimension size is 100. Inputting the global click probability p as a feature into a multi-layer perceptron g (u i ,v j )=0.8345。
Step 4.3: the click probability p calculated in the steps 4.1 and 4.2 l (u i ,v j ) And p g (u i ,v j ) And the weight lambda is calculated through a door mechanism to carry out aggregation to obtain a user u i For candidate item v i Click probability Pr (u) i ,v j )=λp l (u i ,v j )+(1-λ)p g (u i ,v j ) =0.8793. Establishing a corresponding target function, training by using an Adam optimizer to minimize a loss function, and setting the learning rate to be 0.001;
step 4.4: and after one training is finished, updating the behavior and site node potential characteristics, and circulating the calculation processes in the steps 3, 4 and 5 to carry out the next training and result output.
As shown in fig. 6 (a) and 6 (b), the experimental results of this example are as follows:
the Accuracy (AUC) of the test set stabilized at 0.9218. In addition, mean absolute error value (MAE) stabilized at 0.1965, root Mean Square Error (RMSE) stabilized at 0.3361, and F1-score stabilized at 0.8836.
The experimental result shows that the mixed recommendation algorithm based on the joint learning of the local and global representation models can effectively mine useful characteristics in local information and global information related to the project, further can accurately model preference information of the user, and has the advantages of high accuracy of the prediction result, small error and high practical value.
The present invention is not limited to the above-mentioned embodiments, and based on the technical solutions disclosed in the present invention, those skilled in the art can make some substitutions and modifications to some technical features without creative efforts according to the disclosed technical contents, and these substitutions and modifications are all within the protection scope of the present invention.
Claims (7)
1. A hybrid recommendation method based on joint learning of local and global representation models is characterized by comprising the following steps:
step 1: by adopting entity link, the project is linked with the entity in the external knowledge map, and text description information related to the project is obtained and used as local information and an information map of heterogeneous relationship between the entities and used as global information;
and 2, step: constructing a knowledge complement embedded network as a local model according to the local information;
obtaining a local embedded representation of the item by using a knowledge enhanced embedding algorithm on the textual description content of the item; calculating the weight value to represent the correlation degree between two projects, constructing a historical interactive binary graph structure according to the interactive behavior information between the user and the projects, and calculating by combining the two to obtain the local embedded representation of the user;
and 3, step 3: constructing a knowledge perception graph convolutional neural network as a global model according to global information;
and 4, step 4: considering the historical interactive behavior of the user, simulating the local preference and the global preference of the user to the item under different models by combining an attention mechanism, respectively integrating the representation information of the user and the item in local and global embedding spaces, inputting the integrated result as a characteristic into a multi-layer perceptron for prediction, outputting a required result, and finishing a joint learning process;
the step 4 specifically comprises the following steps:
step 4.1: user u i And candidate item v i The local embedding expresses that the cosine similarity is calculated to obtain the local click probability p l ;
Step 4.2, user u i And candidate item v i The global embedding expresses that the cosine similarity is calculated to obtain the global click probability p g ;
Step 4.3: the local click probability p calculated in the steps 4.1 and 4.2 l And global click probability p g And the weight is calculated through a door mechanism to carry out aggregation to obtain a user u i For candidate item v i Establishing a corresponding target function according to the click probability, and training by using an optimizer to minimize a loss function;
step 4.4: after one-time training is finished, the potential characteristics of the behavior and place nodes are updated;
and (5) circulating the calculation processes in the steps 4.3 and 4.4 to carry out next training and result output.
2. The local and global representation model joint learning-based hybrid recommendation method according to claim 1, wherein in the step 1, the names of the items are used as queries and the type information of the items is used as constraints to link the items with the entities in the external knowledge graph, so as to obtain a textual description information related to the items and an information graph of heterogeneous relationships between the entities.
3. The hybrid recommendation method based on the local and global representation model joint learning of claim 2, wherein obtaining the textual description information related to the item and the heterogeneous relationship information graph between the entities comprises:
step 1.1: the method comprises the steps that interactive information between a user and a project and basic information of the project are obtained in a centralized mode in project recommendation data;
step 1.2: respectively representing users in the data set, items in the data set, the number of the items and type information of the items by using a list;
step 1.3: taking the title of the project as query content, searching API by using an offline knowledge graph, and retrieving KB entity from the knowledge graph Freebase;
step 1.4: based on the linked entity, text description information is extracted from the KB, description information corresponding to the item is represented by a list, structure information is extracted, and representation is represented by a heterogeneous graph.
4. The hybrid recommendation method based on joint learning of local and global representation models according to claim 3, characterized in that said data comprises historical interaction data between users and items; the project information comprises a title and a category; the local information and the global information respectively comprise entity description information and heterogeneous relation information in the knowledge graph.
5. The hybrid recommendation method based on joint learning of local and global representation models according to claim 1, wherein obtaining the locally embedded representation of the item using knowledge enhanced embedding algorithm for the item text description content comprises:
step 2.1: aligning the title of the item with the text description information obtained in step 1.4; obtaining each word and the initial embedded representation thereof;
step 2.2: calculating a title-embedded representation of the obtained title and text description using a knowledge-enhanced embedding algorithmAnd textual description sentence embedding representationFinally, the embedded representations of the items are merged to obtain the local embedded representation of the items.
6. The hybrid recommendation method based on the local and global representation model joint learning of claim 1, wherein a historical interactive binary graph structure is constructed, and a user local embedded representation is obtained by combining the historical interactive binary graph structure and the global representation model joint learning, and the method comprises the following steps:
step 2.3: representing user u using lists based on locally embedded representations of items i Setting a specific weight calculation mode for the historical interactive items and the representing candidate items, and respectively calculating the weights between the historical interactive items and the candidate items;
step 2.4: and corresponding the weight to the historical interactive item to represent the contribution degree to modeling user preference information, and combining the weight with the local embedded representation to aggregate to obtain a new embedded representation which indicates the local embedded representation of the user.
7. The hybrid recommendation method based on joint learning of local and global representation models according to claim 1, wherein the step 3, obtaining global embedded representations of items and users by using knowledge-graph convolutional neural network for structural information in knowledge-graph, comprises:
step 3.1: combining the user, the project and the interaction information Y between the user and the project to add the user, the project and the interaction information Y into the abnormal graph H obtained in the step 1.4 to obtain a unified relational graph G;
step 3.2: using an embedded learning algorithm based on transfer to the relationship graph G constructed in the step 3.1 to obtain embedded representation of nodes and relationships in the relationship graph G;
step 3.3: taking a user node in the relational graph G as a central node, and taking a directly connected node as a neighbor node; neighbor node characteristics in the relational graph G are aggregated by adopting a knowledge-aware graph convolution neural network to obtain neighbor embedding expression, and new embedding is obtained by combining with embedding information of a user and is used as the user u i Is embedded in the information.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110026008.5A CN112733027B (en) | 2021-01-08 | 2021-01-08 | Hybrid recommendation method based on local and global representation model joint learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110026008.5A CN112733027B (en) | 2021-01-08 | 2021-01-08 | Hybrid recommendation method based on local and global representation model joint learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112733027A CN112733027A (en) | 2021-04-30 |
CN112733027B true CN112733027B (en) | 2022-10-25 |
Family
ID=75589913
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110026008.5A Active CN112733027B (en) | 2021-01-08 | 2021-01-08 | Hybrid recommendation method based on local and global representation model joint learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112733027B (en) |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113268670B (en) * | 2021-06-16 | 2022-09-27 | 中移(杭州)信息技术有限公司 | Latent factor hybrid recommendation method, device, equipment and computer storage medium |
CN113780002B (en) * | 2021-08-13 | 2023-11-28 | 北京信息科技大学 | Knowledge reasoning method and device based on graph representation learning and deep reinforcement learning |
CN114117220A (en) * | 2021-11-26 | 2022-03-01 | 东北大学 | Deep reinforcement learning interactive recommendation system and method based on knowledge enhancement |
CN114881329B (en) * | 2022-05-09 | 2023-04-07 | 山东大学 | Tire quality prediction method and system based on guide map convolution neural network |
CN115455258B (en) * | 2022-09-14 | 2023-04-18 | 中国电子科技集团公司第十五研究所 | Network space language description and analysis method and device |
CN116662676A (en) * | 2023-06-09 | 2023-08-29 | 北京华品博睿网络技术有限公司 | Online recruitment bidirectional reciprocity recommendation system and method based on multi-behavior modeling |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110717098A (en) * | 2019-09-20 | 2020-01-21 | 中国科学院自动化研究所 | Meta-path-based context-aware user modeling method and sequence recommendation method |
CN111522962A (en) * | 2020-04-09 | 2020-08-11 | 苏州大学 | Sequence recommendation method and device and computer-readable storage medium |
CN112035746A (en) * | 2020-09-01 | 2020-12-04 | 湖南大学 | Session recommendation method based on space-time sequence diagram convolutional network |
CN112085559A (en) * | 2020-08-18 | 2020-12-15 | 山东大学 | Interpretable commodity recommendation method and system based on time-sequence knowledge graph |
EP3752973A1 (en) * | 2018-06-21 | 2020-12-23 | Samsung Electronics Co., Ltd. | System and method for generating aspect-enhanced explainable description-based recommendations |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10643120B2 (en) * | 2016-11-15 | 2020-05-05 | International Business Machines Corporation | Joint learning of local and global features for entity linking via neural networks |
-
2021
- 2021-01-08 CN CN202110026008.5A patent/CN112733027B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP3752973A1 (en) * | 2018-06-21 | 2020-12-23 | Samsung Electronics Co., Ltd. | System and method for generating aspect-enhanced explainable description-based recommendations |
CN110717098A (en) * | 2019-09-20 | 2020-01-21 | 中国科学院自动化研究所 | Meta-path-based context-aware user modeling method and sequence recommendation method |
CN111522962A (en) * | 2020-04-09 | 2020-08-11 | 苏州大学 | Sequence recommendation method and device and computer-readable storage medium |
CN112085559A (en) * | 2020-08-18 | 2020-12-15 | 山东大学 | Interpretable commodity recommendation method and system based on time-sequence knowledge graph |
CN112035746A (en) * | 2020-09-01 | 2020-12-04 | 湖南大学 | Session recommendation method based on space-time sequence diagram convolutional network |
Non-Patent Citations (2)
Title |
---|
A Duet Recommendation Algorithm Based on Jointly Local and Global Representation Learning;xiaoming liu等;《arXiv》;20201202;全文 * |
一种融合知识图谱与长短期偏好的下一项推荐算法;沈冬东等;《小型微型计算机***》;20200409(第04期);第849-854页 * |
Also Published As
Publication number | Publication date |
---|---|
CN112733027A (en) | 2021-04-30 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112733027B (en) | Hybrid recommendation method based on local and global representation model joint learning | |
WO2021179640A1 (en) | Graph model-based short video recommendation method, intelligent terminal and storage medium | |
US11714831B2 (en) | Data processing and classification | |
CN109389151B (en) | Knowledge graph processing method and device based on semi-supervised embedded representation model | |
US8566321B2 (en) | Relativistic concept measuring system for data clustering | |
CN114048331A (en) | Knowledge graph recommendation method and system based on improved KGAT model | |
CN112836120A (en) | Multi-mode knowledge graph-based movie recommendation method, system and terminal | |
Wen et al. | Neural attention model for recommendation based on factorization machines | |
WO2023093205A1 (en) | Entity tag association prediction method and device and computer readable storage medium | |
CN112884551B (en) | Commodity recommendation method based on neighbor users and comment information | |
CN112800344B (en) | Deep neural network-based movie recommendation method | |
Yang et al. | A decision-making algorithm combining the aspect-based sentiment analysis and intuitionistic fuzzy-VIKOR for online hotel reservation | |
WO2023179689A1 (en) | Knowledge graph-based recommendation method for internet of things | |
Wang et al. | Trust-aware collaborative filtering with a denoising autoencoder | |
CN113326384A (en) | Construction method of interpretable recommendation model based on knowledge graph | |
WO2023284516A1 (en) | Information recommendation method and apparatus based on knowledge graph, and device, medium, and product | |
CN114386513A (en) | Interactive grading prediction method and system integrating comment and grading | |
CN112257959A (en) | User risk prediction method and device, electronic equipment and storage medium | |
Ma et al. | GAFM: A knowledge graph completion method based on graph attention faded mechanism | |
Lee | Extraction of competitive factors in a competitor analysis using an explainable neural network | |
Ren et al. | A co-attention based multi-modal fusion network for review helpfulness prediction | |
Yu et al. | A graph attention network under probabilistic linguistic environment based on Bi-LSTM applied to film classification | |
Huang et al. | Multi-affect (ed): improving recommendation with similarity-enhanced user reliability and influence propagation | |
CN114168804B (en) | Similar information retrieval method and system based on heterogeneous subgraph neural network | |
CN114238752B (en) | Article recommendation method, device and storage medium |
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 |