CN115953215A - Search type recommendation method based on time and graph structure - Google Patents
Search type recommendation method based on time and graph structure Download PDFInfo
- Publication number
- CN115953215A CN115953215A CN202211533857.0A CN202211533857A CN115953215A CN 115953215 A CN115953215 A CN 115953215A CN 202211533857 A CN202211533857 A CN 202211533857A CN 115953215 A CN115953215 A CN 115953215A
- Authority
- CN
- China
- Prior art keywords
- user
- item
- search
- history
- users
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 59
- 239000013598 vector Substances 0.000 claims abstract description 92
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 71
- 230000006399 behavior Effects 0.000 claims abstract description 24
- 238000013528 artificial neural network Methods 0.000 claims abstract description 21
- 230000008569 process Effects 0.000 claims abstract description 19
- 230000008447 perception Effects 0.000 claims abstract description 11
- 238000005070 sampling Methods 0.000 claims abstract description 7
- 230000006870 function Effects 0.000 claims description 27
- 238000004364 calculation method Methods 0.000 claims description 17
- 238000002474 experimental method Methods 0.000 claims description 17
- 238000005295 random walk Methods 0.000 claims description 13
- 238000012360 testing method Methods 0.000 claims description 13
- 230000007246 mechanism Effects 0.000 claims description 12
- 238000012549 training Methods 0.000 claims description 12
- 230000003044 adaptive effect Effects 0.000 claims description 11
- 230000002776 aggregation Effects 0.000 claims description 8
- 238000004220 aggregation Methods 0.000 claims description 8
- 238000010845 search algorithm Methods 0.000 claims description 8
- 230000007423 decrease Effects 0.000 claims description 3
- 238000012886 linear function Methods 0.000 claims 1
- 238000013507 mapping Methods 0.000 abstract description 2
- 230000000694 effects Effects 0.000 description 8
- 239000000047 product Substances 0.000 description 8
- 230000007774 longterm Effects 0.000 description 6
- 238000012512 characterization method Methods 0.000 description 5
- 238000011156 evaluation Methods 0.000 description 4
- 238000012952 Resampling Methods 0.000 description 3
- 238000013459 approach Methods 0.000 description 3
- 230000003542 behavioural effect Effects 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000000052 comparative effect Effects 0.000 description 2
- 230000002708 enhancing effect Effects 0.000 description 2
- 230000001537 neural effect Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 238000007430 reference method Methods 0.000 description 2
- 230000004931 aggregating effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000007876 drug discovery Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000002787 reinforcement Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a search type recommendation method based on time and a graph structure, which relates to the field of recommendation systems and comprises the following steps: collecting user behavior historical data of an internet platform, coding users and articles in the historical data by using an inner product neural network, calculating embedded vectors of the users and the articles, mapping, sampling and learning the historical articles, inputting the embedded vectors of the articles into a model, searching similar articles and similar users in the user historical data based on the characteristics of the users and the articles, and searching similar articles of a target article from the historical data of the similar users; inputting the retrieved user history sequence into a time perception model to obtain a hidden state sequence, inputting the hidden state sequence into a multilayer perceptron to predict and perform back propagation, and applying the trained algorithm model to a recommendation algorithm. The recommendation method provided by the invention can process large-scale user history sequences in an online environment, and can effectively improve the recommendation efficiency and performance.
Description
Technical Field
The invention relates to the field of recommendation systems, in particular to a search type recommendation method based on time and graph structures.
Background
With the rapid increase of the history length of user behaviors, how to effectively help users find out the commodities which are interested in the users from the huge number of candidate commodities on the internet platform has become a very important problem in the field of recommendation systems. Classical recommendation methods, such as collaborative filtering models and factorization models, mainly focus on modeling the overall interest of a user in order to obtain items of interest to the user. But such models explore less the user's needs over a fixed period of time, which is in fact one of the very important factors affecting the user's needs. Therefore, there are some recent attempts to capture the user's sequence pattern by means of a memory network, a recurrent neural network or a time point model, but most methods cannot be applied in a practical internet environment due to the computational complexity and the limitation of storage space when the user's historical sequence is too long.
Classical recommendation algorithms typically recommend an item that a user wants through a rich user-item interaction history, which is typically in a table, sequence, or graph form. However, according to the user behavior search for click rate prediction published in the international information search conference by Qin Jiarui and the like, as user behavior data and the like are accumulated more and more, it is very difficult to train a model from the whole user log due to the limitation of online calculation. One possible approach is to focus only on the user's recent history, using short-term history instead of long-term history, to generate personalized recommendations. Recent papers, however, indicate that (e.g., "search-based user interest modeling and long-term sequential behavior for click-through rate prediction" at information and knowledge management meetings), these approaches fail to encode the user's demand dependence on periodicity and long-term, so focusing only on the user's recent history is actually a sub-optimal solution. Both papers then suggest that both hard and soft searches can be performed across the entire user history. Qin Jiarui also proposes modeling query construction for context data by reinforcement learning and then using this module to query retrieve relevant behavior using BM25 relevance functions. There is also a paper (search-based user interest modeling and long-term sequential behavior for click-through rate prediction) that proposes matching related items by hard retrieval of item classes and soft retrieval of matching related items based on item characterization vectors. In this manner, the model may use the relevant items retrieved throughout the user's behavioral history to make recommendations. But existing search-based methods ignore time intervals in the user's behavioral history. And generally the retrieved related items are all positively fed back (e.g., the user clicks on the item), which gives the opportunity to consider both positively and negatively fed back related items so that the entire user sequence is fully utilized.
Articles such as "modeling user behavior with time LSTM model" published in artificial intelligence international federation meetings and "recursive poisson factorization of time recommendations" Hou Saini published in the journal of knowledge and data engineering, etc., capture the user's performance better by using the time interval between user behaviors that conventional sequence structures cannot use. One direction is to control the updates of short-term and long-term interest by using new dedicated time gates, such as Zhao Pengpeng a "space-time LSTM model for recommending next points of interest" proposes a distance gate based on a recurrent neural network to control the updates of short-term and long-term points of interest. Another method of using time interval information is to specify the user's sequence history by a point process, i.e. discrete times in the user history are modeled as continuous times, such as Mei Hongyuan, etc. published in the neural information processing system congress paper "neural hox process: a neuro-self-regulating multivariable point process "proposes a hokes process that allows past events to influence future predictions through a complex and realistic style. In addition to the high computational complexity and time consumption of these methods, the direct input of long user sequences into the model also introduces very large noise (interference information), which makes it unfeasible to capture rich sequence patterns directly in the user log.
In the fields of recommendation systems, social networks, drug discovery, mathematical programming and the like, a characterization vector obtained by learning nodes on a graph is used as a basic data module in a model, so that the deep learning method has great application. Mainstream graph characterization learning is primarily to establish more generalized node proximity relations by modeling the adjacency matrix of the graph or by random walk and fitting the actual proximity relations. The most common deep walking model is one of the earliest models using random walking, and a sufficient number of node sequences are obtained by random walking in a directed graph and are input into a jump graph model to learn the embedded vectors of nodes by regarding the node sequences as sentences similar to natural language. The node vector model is a variant of the depth walk model, which controls the tendency of random walk to become biased random walk by two parameters, p and q. In addition, an additional information enhanced graph representation learning model is provided, wherein the model builds a node graph of a building through continuous article access behaviors of hundreds of millions of users on the Internet, samples node sequences through random walks, and trains an embedded vector of nodes through a hopping graph model with characteristics. Wang Hongwei "using a generative confrontation net to learn a graph characterization" published in knowledge and data engineering journal, which proposes a model for learning a graph characterization through a confrontation generation network, the model firstly performs strategy random walk and restart on a graph based on a learnable graph operator, samples node pairs of a central node and a last node before restart, and then a discriminator performs learning to judge whether a pair of nodes has an edge. The embedded vectors of nodes are learned by this countermeasure approach.
For the summary of relevant research at home and abroad: the existing widely used recommendation algorithm in the industry is difficult to solve the problem that online calculation is too complex when the user behavior history length is too long, and meanwhile, the common algorithm cannot effectively utilize the user behavior history, so that the problems of neglecting the time interval of the user behavior, negative feedback of the user and the like exist. In addition, the time perception algorithm in recent years is generally difficult to solve the problem that the direct input time interval brings larger noise and the like.
Therefore, those skilled in the art are devoted to developing a more efficient model for user historical behavior and reasonable time perception, which can perform better than the conventional recommendation algorithm in more situations.
Disclosure of Invention
In view of the above-mentioned drawbacks of the prior art, an object of the present invention is to provide a search-based prediction model that combines feedback information on an item in user history data with historical sequence information to improve the performance of the model in predicting a long-tailed user.
In order to achieve the above object, the present invention provides a search-based recommendation method based on time and graph structure, which is characterized in that the method comprises the following steps:
s101: collecting historical data of internet platform user behaviors, coding the user and the articles in the historical data by using an inner product neural network, and calculating an embedded vector of the user and an embedded vector of the articles by using the inner product neural network;
s103: drawing, sampling and learning historical articles of the user, capturing the associated information among the articles, and inputting the embedded vectors of the articles into a model as article features;
s105: based on the user and the item features, using a search model to retrieve similar items of a target item in the user history data;
s107: acquiring similar users of the target user by using the search model, and retrieving similar items of the target item from historical data of the similar users;
s109: inputting the retrieved user history sequence into a time perception model, and obtaining a hidden state sequence after passing through a gate cycle unit and an attention mechanism, wherein the user history sequence comprises user history feedback information and a user history time interval;
s111: inputting the hidden state sequence into a multilayer perceptron for prediction, and performing back propagation according to a first loss function;
s113: and applying the trained algorithm model to a recommendation algorithm.
Further, in step S101, when the inner product neural network is used to calculate the embedding vector, the following method is adopted:
wherein ,
an embedded vector representing the nth item>
An attribute vector representing the nth item>
and />
Representing a weight in the calculation process>
Representing the pth item in the attribute vector for the nth item.
Further, in step S103, when the embedding vector of the item is obtained, an expansion algorithm for enhancing graph characteristic learning is adopted, and an additional information enhanced graph embedding model is used to model the item sequence in the user history.
Further, when the historical item of the user is mapped in step S103, a positive sample and a negative sample are sampled in the map, the positive sample is resampled by random walk, the item type is used as additional information, and the positive sample, the negative sample and the additional information are input to the additional information enhanced map embedding model for learning, so as to obtain an embedding vector of the item.
Further, the random walk is a walking mode in node embedding learning, training is performed using the second loss function after the positive sample and the negative sample are sampled, and the article embedding vector is calculated by adopting the following formula:
the second loss function is:
wherein ,
an embedded vector representing the nth item>
Is a weight parameter, is>
Represents i m A negative sample of the article.
Further, in step S105, the search model uses an adaptive search algorithm, the adaptive search algorithm supports a hard search and a soft search, the hard search is a search in which the search item types are completely the same, the soft search is a search in which the search item types are completely dissimilar, and the adaptive search algorithm calculates the similarity between the target item type and the user history item type in the following manner:
wherein ,
item set H representing a similar category to the target item in the history of the mth user m Represents the history of the mth user>
Representing recent history in the mth user, i n An identification number representing the nth item>
Indicates the type of the nth target item>
Type of search item, x, indicating nth object item i An attribute vector representing an item, wherein>
A pth item in the attribute vector representing the nth item; tau is used to control the degree of similarity, the greater tau, the closer tau is to similarity, and tau decreases from large to small during training.
Further, in step S107, the similarity between users is calculated by calculating the same kind of articles in the history of the target user and other users, and then the similar user of the target user is retrieved, the history behavior of the similar user is supplemented as additional information, and the similarity between users is calculated as follows:
wherein ,
represents a set of users similar to the mth user, U represents the total users, H m Represents the history of the mth user, u m An identification number representing the mth user, in conjunction with a subscriber number associated with the mth subscriber>
Indicates the type of the mth target user, and>
indicates the type of the retrieving user of the mth target user, based on the status of the retrieving user>
And an embedded vector representing the mth user, wherein the iota is used for controlling the similarity degree, the larger the iota is, the closer the iota is to the similarity, and the iota is greatly reduced in the training process.
Further, in step S109, a preliminary result of the hidden state is obtained through calculation by the gate cycle unit, the preliminary result is further obtained through calculation by the attention mechanism, and the hidden state set is aggregated through an aggregation function;
wherein, the gate cycle unit adopts the following calculation formula:
h′ t ＝f′ t ⊙c′ t +(1-f′ t )⊙h′ t-1 ，
the attention mechanism adopts the following calculation formula:
f t ＝α′ t ·f′ t
h t ＝f t ⊙c t +(1-f t )⊙h′ t ，
the aggregation function is:
wherein ,
characteristic vector representing the t-th item>
A user feedback vector representing the user's for the t-th item,
is->
and />
Vector of spliced, h' t and h′t-1 Is a preliminary hidden state, h t Is in a hidden state, W x and Ux For the weighting parameter, Δ t is two items i t-1 and it The time interval between, de (-) is a heuristic decay algorithm, taking de (Δ t) =1/Δ t when the time interval is short, taking de (Δ t) =1/log (e + Δ t) when the time interval is long,
Further, in step S111, the multi-layer perceptron is a neural network combining an intelligent perceptron and a nonlinear function, and the neural network is:
the first loss function is:
wherein ,
is a hidden state set of similar users of the current user, is based on>
Further, step S113 further includes testing the recommendation algorithm, where the testing environment of the testing includes recording the testing result based on a public data set and an online experiment, and comparing the result difference between the recommendation method and other models.
In the preferred embodiment of the invention, compared with the prior art, the invention has the following beneficial effects:
1. the sequence recommendation algorithm based on search and time perception provided by the invention is gradually transited from hard search to soft search, so that the search process is more reasonable, compared with the traditional classical sequence recommendation algorithm (including the time perception sequence recommendation algorithm), the method can process large-scale user historical sequences in an online environment, and can effectively improve the recommendation efficiency and performance;
2. the method takes the historical feedback of the user as additional information to be input into the model, so that the model has better effect, models the user-object by a graph characteristic learning method, and captures the association between the objects by learning the embedded vector of the object.
The conception, the specific structure and the technical effects of the present invention will be further described with reference to the accompanying drawings to fully understand the objects, the features and the effects of the present invention.
Drawings
FIG. 1 is a schematic flow chart of a preferred embodiment of the present invention;
FIG. 2 is a comparison of experimental results on an offline public data set of a preferred embodiment of the present invention;
FIG. 3 is a comparison experiment result on an online scene for a preferred embodiment of the present invention;
FIG. 4 is a comparison experimental result of the algorithm herein for two configurations on an online scene according to a preferred embodiment of the present invention;
FIG. 5 is a diagram illustrating the results of the statistical comparison between the expansion algorithm and the original algorithm in buckets on-line scene according to a preferred embodiment of the present invention.
Detailed Description
The technical contents of the preferred embodiments of the present invention will be made clear and easily understood by referring to the drawings attached to the specification. The present invention may be embodied in many different forms of embodiments and the scope of the invention is not limited to the embodiments set forth herein.
In the drawings, elements that are structurally identical are represented by like reference numerals, and elements that are structurally or functionally similar in each instance are represented by like reference numerals. The size and thickness of each component shown in the drawings are arbitrarily illustrated, and the present invention is not limited to the size and thickness of each component. The thickness of the components may be exaggerated where appropriate in the figures to improve clarity.
As shown in fig. 1, the search-based recommendation method based on time and graph structure provided in the embodiment of the present invention includes the following steps:
s101: collecting historical data of internet platform user behaviors, coding the user and articles in the historical data by using an inner product neural network, and calculating embedded vectors of the user and the articles by using the inner product neural network.
When the inner product neural network is used for calculating the embedded vector, the following method is adopted:
wherein ,
an embedded vector representing the Nth item, <' >>
An attribute vector representing the nth item>
and />
Represents a weight in the calculation process, and>
representing the pth item in the attribute vector for the nth item.
S103: the method comprises the steps of mapping, sampling and learning historical articles of a user, capturing correlation information among the articles, and inputting embedded vectors of the articles into a model together as article features.
When the embedded vector of the object is obtained, an expansion algorithm for enhancing by graph characteristic learning is adopted, and the object sequence in the user history is modeled by using an additional information enhanced graph embedded model. When the historical articles of the user are mapped, sampling positive samples and negative samples in the map, resampling the positive samples through random walk, using the article types as additional information, inputting the positive samples, the negative samples and the additional information into an additional information enhancement map embedding model for learning, and obtaining the embedding vectors of the articles. When the positive sample is resampled by random walk, the random walk is a walk mode in node embedding learning, the second loss function is used for training after the positive sample and the negative sample are sampled, and an article embedding vector is calculated by adopting the following formula:
the second loss function is:
wherein ,
an embedded vector representing the nth item>
Is a learnable weight parameter, <' > is>
Represents i m A negative sample of the article.
S105: similar items for the target item are retrieved in the user history data using a search model based on the user and item characteristics.
Items similar to the target item are retrieved from the user history by calculating a similarity between the category of the target item and the category of the items of the user history. The calculation of the similarity of the articles is self-adaptive, and the hard similarity (which requires the articles to be completely the same in extreme cases) which requires high is gradually transited to the soft similarity (which requires the articles to be completely dissimilar in extreme cases) along with the training process. The search model described above uses an adaptive search algorithm that supports both hard and soft searches, hard search being a search in which the types of search items are completely the same, soft search being a search in which the types of search items are completely dissimilar, the adaptive search algorithm calculating the similarity of the target item type and the user history item type as follows:
wherein ,
item set H representing a similar category to the target item in the history of the mth user m Represents the history of the mth user>
Representing recent history in the mth user, i n Indicates the identification number, which stands for the nth item>
Indicates the type of the nth target item>
Type, x, of search item representing nth object item i An attribute vector representing an item, wherein>
A pth item in the attribute vector representing the nth item; tau is used to control the degree of similarity, the greater tau, the closer tau is to similarity, and tau decreases from large to small during training.
S107: similar users of the target user are obtained by using the search model, and similar items of the target items are retrieved from historical data of the similar users.
Similarity between the users is calculated by calculating the same kind of articles in the histories of the target user and other users, so that similar users of the target user are retrieved, and the historical behaviors of the similar users are supplemented as additional information.
The similarity between users is calculated as follows:
wherein ,
represents a set of users similar to the mth user, U represents the total users, H m Represents the history of the mth user, u m An identification number representing the mth user, and->
Indicates the kind of the mth target user>
Indicates the type of the retrieving user of the mth target user, based on the status of the retrieving user>
And an embedded vector representing the mth user, wherein the iota is used for controlling the similarity degree, the larger the iota is, the closer the iota is to the similarity, and the iota is greatly reduced in the training process.
S109: and inputting the retrieved user history sequence into a time perception model, and obtaining a hidden state sequence after passing through a gate cycle unit and an attention mechanism, wherein the user history sequence comprises user history feedback information, user history time interval and other information.
And calculating by the gate cycle unit to obtain a preliminary result of the hidden state, further calculating by an attention mechanism to obtain a final hidden state, and aggregating the hidden state set by an aggregation function.
Wherein, the door cycle unit adopts the following calculation formula:
h′ t ＝f′ t ⊙c′ t +(1-f′ t )⊙h′ t-1 ，
the attention mechanism adopts the following calculation formula:
f t ＝α′ t ·f′ t
h t ＝f t ⊙c t +(1-f t )⊙h′ t ，
the aggregation function is:
wherein ,
represents the characteristic vector of the tth item>
A user feedback vector representing the user's contribution to the tth item,
is->
and />
Vector of spliced, h' t and h′t-1 Is a preliminary hidden state, h t Is in a hidden state, W x and Ux For the weighting parameter, Δ t is two items i t-1 and it The time interval between, de (-) is a heuristic decay algorithm, taking de (Δ t) =1/Δ t when the time interval is short, taking de (Δ t) =1/log (e + Δ t) when the time interval is long,
S111: and inputting the hidden state sequence into a multilayer perceptron to predict, and performing back propagation according to a first loss function.
The multilayer perceptron is a neural network combining an intelligent perceptron and a nonlinear function, and the neural network is as follows:
the first loss function is:
wherein ,
is a hidden state set of similar users of the current user, is based on>
S113: and applying the trained algorithm model to a recommendation algorithm.
And testing the recommendation algorithm, wherein the testing comprises recording a test result based on a public data set and an online experiment, and comparing the result difference of the recommendation method and other models.
The search type recommendation method based on the time and the graph structure provided by the embodiment of the invention has the following technical effects:
1. the invention provides a sequence recommendation algorithm based on search and time perception, and compared with the traditional classical sequence recommendation algorithm (including the time perception sequence recommendation algorithm), the sequence recommendation algorithm can process large-scale user history sequences in an online environment.
2. The invention provides a self-adaptive search-based algorithm, which gradually transits from hard search to soft search, so that the search process is more reasonable.
3. According to the invention, the historical feedback of the user is input into the model as additional information, and the model using the skill has a better experiment effect.
4. In the invention, a graph characteristic learning method is used for modeling the user-object, and the association between the objects is captured by learning the embedded vector of the objects.
5. The method can obtain better effect compared with the domestic and foreign advanced recommendation algorithm on three types of classical data sets from real internet application, and can effectively improve the recommendation efficiency and performance in an online recommendation scene.
As shown in fig. 2 to 5, for the search-based recommendation method based on time and graph structure provided in the embodiment of the present invention, in order to verify the experimental results of the present invention, the present invention shows comparative experimental results in three types of real offline public data sets and a real online recommendation scene, and for the offline public data sets, the comparative baseline algorithm is 3 reference methods commonly used in the recommendation algorithm: deep interest network, deep interest evolution network and behavior modeling based on search three recommendation algorithms.
The comparison test is respectively carried out on three public data sets of Tianmao, paibao and Taobao and a recommendation scene of 'guessing you like' of a tenderer bank, the first three data sets are from application of a real internet platform, and the recommendation scene of 'guessing you like' is an online scene which is very important in the industry. In an off-line data set comparison experiment, main evaluation indexes are the area under the ROC curve (AUC), the Accuracy (ACC) and the logarithmic loss value (Logloss), "the text algorithm-" is the configuration without using the label skill in the algorithm, and "the text algorithm +" is the configuration using the graph to characterize and learn in the algorithm; for the online recommendation scene of 'guessing you like', the main evaluation indexes are Click Through Rate (CTR), AUC, average item click rate (AIC) and user click rate (CUR), when an expansion algorithm and an original algorithm are compared, average click times (CPC) are also used, wherein indexes with w/o subscripts are results obtained after users without clicks are removed, bucket-dividing statistics is carried out on the expansion experiment, the experiment results of the users with different history lengths under the two algorithms are compared in a statistical mode (the difference value of CTR refers to the CTR result of the expansion algorithm minus the CTR value of the original algorithm), and the overall comparison experiment effect is shown in 1,2,3,4. It can be observed from the figure that the result of the invention compared with the result of the baseline recommendation algorithm can obtain better result on the benefit index, which shows that the technical scheme of the invention is more effective, and meanwhile, the expanded algorithm has better result compared with the original algorithm.
The present invention will be described in detail below with reference to preferred embodiments thereof.
As shown in fig. 1, the implementation process of the present invention includes the following steps:
step one, collecting real user behavior history on an internet platform, and coding the types of the user and the object by using an inner product neural network for unified processing.
In this step, the basic elements are defined as follows:
(1) user identification number u and item identification number i: a vector representing an identification number unique to each different user and item, where u m Identification number, i, representing mth user n An identification number representing an nth item;
(2) user attribute vector x u And an item attribute vector x i : an attribute vector representing the user and the item, wherein
Represents the pth entry in an attribute vector for the mth user, and { }>
A pth item in the attribute vector representing the nth item;
(3) user embedding vector e u And an item embedding vector e i : an embedded vector representing the user and the item, wherein
An embedded vector representing the mth user, and->
An embedded vector representing the nth item;
the inner product neural network is used on this basis to compute the embedding vector:
wherein 
and />
Representing a weight learnable in the calculation process>
The calculation method of (1) is the same as above.
And step two, using a graph characteristic learning to carry out an enhanced expansion algorithm, carrying out graph building, sampling and learning on historical articles of a user, capturing the associated information among the articles, and inputting the learned articles embedded vectors serving as article characteristics into the model together.
The sequence of items in the user history is modeled using an additional information enhancement graph embedding model, obtaining an embedding vector for its items. The method comprises the steps of establishing a graph of an article in user history, sampling positive and negative samples in the graph, resampling the positive sample through random walk, using the article type as additional information, learning the obtained positive sample, negative sample and additional information through an additional information enhancement graph embedding model, obtaining an embedding vector, and processing the embedding vector as additional features of the article. The walking mode used by resampling is a walking mode in node embedding learning, and the mode of calculating an article embedding vector is as follows:
wherein ,
an embedded vector representing an nth item>
Is a learnable weight parameter, and the second loss function is:
wherein ,
represents i m A negative sample of the article.
And step three, based on the user and article characteristics obtained in the step one and the step two, using a search-based model for the target user and the target article, and using an adaptive algorithm to search for articles with higher similarity to the target article in the history of the user.
And (3) the user and item code pairs obtained in the first step and the second step are used for retrieving items in the user history similar to the target item through an adaptive search-based module. Items similar to the target item are retrieved from the user history by calculating a similarity between the category of the target item and the category of the user history items. The calculation of the similarity of the articles is self-adaptive, and the hard similarity (which requires the articles to be completely the same in extreme cases) which requires high is gradually transited to the soft similarity (which requires the articles to be completely dissimilar in extreme cases) along with the training process. Under the self-adaptive algorithm, the whole algorithm framework forms an end-to-end style. In addition to calculating the similarity between the target item type and the user historical item type, the adaptive algorithm is also applied to find similar users. The similarity between the users is calculated by calculating the objects of the same type in the histories of the target user and other users, so that the users similar to the target user are searched, and the historical behaviors of the users are used as additional information to supplement, so that a better effect is achieved.
On this basis, the basic elements are defined as follows:
(1) item attribute vector x i : an attribute vector representing the item, wherein
A pth item in the attribute vector representing the nth item;
(2) user embedding vector e u : an embedded vector representing a user, wherein
An embedded vector representing the mth user;
(3) class c of user u And article type c i : indicates the kinds of users and articles, wherein
Indicates the kind of the mth target user>
Indicates the type of the retrieving user of the mth target user, based on the status of the retrieving user>
Indicates the type of the nth target item>
A search item type indicating an nth target item;
(4) user identification number u and item identification number i: a vector representing an identification number unique to each different user and item, where u m Identification number, i, representing mth user n Denotes the nth objectAn identification number of the article;
on the basis, searching for similar articles:
wherein ,
item set H representing a similar category to the target item in the history of the mth user m Represents the history of the mth user>
And (3) representing the recent history in the mth user, wherein tau is used for controlling the degrees of hard similarity and soft similarity, the larger tau is, the closer tau is to the hard similarity, and tau is reduced from large to small in the training process.
The method for finding similar users is similar:
wherein ,
representing a set of users similar to the mth user, U representing the total users, iota being used to control the degree of hard similarity and soft similarity, the greater iota being closer to hard similarity, the smaller iota becomes during training.
And step four, similarly to the step three, obtaining users with higher similarity to the target user by using a self-adaptive algorithm, and retrieving the articles with higher similarity to the target articles in the history of the users.
For a given user u m And given target item i n Thereafter, a similar item set can be retrieved through step 3
And a similar user->
Can be written into and/or taken out>
For->
Each article i in n Let us know user u m Feedback (click or not, etc.) on the item. And then inputting the feature vector of the object and the user feedback click splicing into a gate cycle unit for learning to discover a useful sequence pattern. After passing through the gate cycle unit, the change in user interest is modeled using an attention mechanism, and this feature of time intervals is taken into account. The hidden state sequence obtained after the gate cycle unit and the attention mechanism obtains a vector for representing the hidden state sequence through a designed aggregation function. On this basis, the following basic elements are defined:
(1) feature vector of an article
A vector representing features of the tth item;
(2) user feedback vector
Representing an embedding vector obtained according to the feedback of the user to the t-th item;
inside the gate cycle unit, we compute the hidden state by:
h′ t ＝f′ t ⊙c′ t +(1-f′ t )⊙h′ t-1
in this formula, W x and Ux Are all weight parameters that can be learned,
is->
and />
Vector by concatenation, and h' t and h′t-1 Is a preliminary hidden state. After passing through the gate cycle unit, the preliminary hidden state is further calculated by the following attention mechanism:
f t ＝α′ t ·f′ t
h t ＝f t ⊙c t +(1-f t )⊙h′ t
in this formula, W x Is a learnable weight parameter, h t Is a hidden state. Δ t is two items i t-1 and it The time interval between de (-) is a heuristic decay algorithm that takes de (Δ t) =1/Δ t when the time interval is short and de (Δ t) =1/log (e + Δ t) when the time interval is long.
For each set
We can get a series of hidden states->
Each set of hidden states may be aggregated by an aggregation function as follows:
wherein 
For users similar to the current user, the same method is used to calculate ≥ s>
wherein />
And step five, inputting the retrieved user history sequence into a time perception model, and inputting the user history feedback information and the user history time interval into the model as characteristics.
The neural network ultimately used for prediction is a combination of an intelligent perceptron and a nonlinear function, as follows:
the first loss function used therein is:
and step six, inputting the obtained hidden state into a multilayer perceptron for prediction, and performing back propagation according to the first loss function.
And step seven, applying the trained algorithm model to the public data set and an online experiment, recording the experiment result, and comparing the difference between the result and the result of other models.
For the search-type recommendation method based on time and graph structure provided by the embodiment of the invention, comparison experiments are performed in three real off-line public data sets and a real on-line recommendation scene, and the test results are shown in fig. 2-5. During comparison experiments, the off-line public data set selects three applications from a real internet platform, namely tianmao, paibao, taobao and the like, and the online scene selection recruiter bank selects a recommendation scene of ' guessing you like ' (the recommendation scene of guessing you like ' is an online scene which is very important in the industry). For an offline public data set, the baseline algorithm for comparison is 3 reference methods commonly used in the recommendation algorithm: deep interest network, deep interest evolution network and behavior modeling based on search three recommendation algorithms. In an off-line data set comparison experiment, main evaluation indexes are the area under the ROC curve (AUC), the Accuracy (ACC) and the logarithmic loss value (Logloss), "the text algorithm-" is the configuration without using the label skill in the algorithm, and "the text algorithm +" is the configuration using the graph to characterize and learn in the algorithm; for the 'guess you like' online recommendation scene, the main evaluation indexes are Click Through Rate (CTR), AUC, average item click rate (AIC) and user click rate (CUR), when an extended algorithm and an original algorithm are compared, average click times (CPC) are also used, wherein indexes with w/o subscripts are results obtained after users without clicks are removed, barrel counting is carried out on the extended experiment, and the experiment results of the users with different historical lengths under the two algorithms are counted and compared (the difference value of CTR refers to the CTR value of the CTR result of the extended algorithm minus the CTR value of the original algorithm). It can be observed from fig. 2-5 that the results of the present invention can obtain better results in the benefit index than the results of the baseline recommendation algorithm, which indicates that the technical scheme of the present invention is more effective, and meanwhile, the extended algorithm has better results than the original algorithm.
The foregoing detailed description of the preferred embodiments of the invention has been presented. It should be understood that numerous modifications and variations could be devised by those skilled in the art in light of the present teachings without departing from the inventive concepts. Therefore, the technical solutions available to those skilled in the art through logic analysis, reasoning and limited experiments based on the prior art according to the concept of the present invention should be within the scope of protection defined by the claims.
Claims (10)
1. A search-based recommendation method based on time and graph structure is characterized by comprising the following steps:
s101: collecting historical data of internet platform user behaviors, coding the user and an article in the historical data by using an inner product neural network, and calculating an embedded vector of the user and an embedded vector of the article by using the inner product neural network;
s103: drawing, sampling and learning historical articles of the user, capturing the associated information among the articles, and inputting the embedded vectors of the articles into a model as article features;
s105: based on the user and the item features, using a search model to retrieve similar items of a target item in the user history data;
s107: acquiring similar users of the target user by using the search model, and retrieving similar items of the target item from historical data of the similar users;
s109: inputting the retrieved user history sequence into a time perception model, and obtaining a hidden state sequence after a gate cycle unit and an attention mechanism, wherein the user history sequence comprises user history feedback information and a user history time interval;
s111: inputting the hidden state sequence into a multilayer perceptron to predict, and performing back propagation according to a first loss function;
s113: and applying the trained algorithm model to a recommendation algorithm.
2. The method of claim 1, wherein in step S101, when calculating the embedding vector using the inner product neural network, the following is adopted:
wherein ,
an embedded vector representing the nth item>
An attribute vector representing an nth item>
and />
Represents a weight in the calculation process, and>
the pth entry in the attribute vector representing the nth item.
3. The method according to claim 2, wherein in step S103, when obtaining the embedding vector of the item, an expansion algorithm enhanced by graph feature learning is adopted, and an additional information enhanced graph embedding model is used to model the item sequence in the user history.
4. The method according to claim 3, wherein, when the user' S historical articles are mapped in step S103, positive and negative samples are sampled in the map, the positive samples are resampled by random walk, the article type is used as additional information, and the positive samples, the negative samples and the additional information are input into the additional information enhancement map embedding model for learning, so as to obtain the embedding vector of the article.
5. The method of claim 4, wherein the random walk is a walk in node-embedding learning, trained using the second loss function after the positive and negative samples are sampled, the item-embedding vector being calculated using the following formula:
the second loss function is:
wherein ,
an embedded vector representing the nth item>
Is a weight parameter, is->
Represents i m A negative sample of the article.
6. The method according to claim 5, wherein in step S105, the search model uses an adaptive search algorithm, the adaptive search algorithm supports hard search and soft search, the hard search is search for retrieving the same kind of items, the soft search is search for retrieving the different kinds of items, the adaptive search algorithm calculates the similarity between the target kind of items and the user history kind of items as follows:
wherein ,
item set H representing a similar category of the target item in the history of the first user m Represents the history of the mth user>
Representing recent history in the mth user, i n Indicates the identification number, which stands for the nth item>
Indicates the type of the nth target item>
Type, x, of search item representing nth object item i An attribute vector representing an item, wherein>
A pth item in the attribute vector representing the nth item; tau is used to control the degree of similarity, the greater tau, the closer tau is to similarity, and tau decreases from large to small during training.
7. The method according to claim 6, wherein in step S107, similarity between users is calculated by calculating the same kind of articles in the history of the target user and other users, and then similar users of the target user are retrieved, the history behavior of the similar users is supplemented as additional information, and the similarity between users is calculated as follows:
wherein ,
represents a set of users similar to the mth user, U represents the total users, H m Represents the history of the mth user, u m An identification number representing the mth user, and->
Indicates the kind of the mth target user>
Indicates the type of the retrieving user of the mth target user, based on the status of the retrieving user>
The embedded vector representing the mth user, iota is used to control similarTo a greater extent, the greater the iota, the closer the similarity, and the smaller iota becomes during training.
8. The method according to claim 7, wherein in step S109, a preliminary result of the hidden states is obtained by the gate loop unit, and the preliminary result is further obtained by the attention mechanism, and the set of hidden states is aggregated by an aggregation function;
wherein, the gate cycle unit adopts the following calculation formula:
h′ t ＝f′ t ⊙c′ t +(1-f′ t )⊙h′ t-1 ，
the attention mechanism adopts the following calculation formula:
f t ＝α′ t ·f′ t
h t ＝f t ⊙c t +(1-f t )⊙h′ t ，
the aggregation function is:
wherein ,
represents the characteristic vector of the tth item>
Represents a user feedback vector, based on the user's status of the tth item>
Is->
and />
Vector of spliced, h' t and h′t-1 Is a preliminary hidden state, h t Is in a hidden state, W x and Ux For the weighting parameter, Δ t is two items i t-1 and it The time interval between, de (-) is a heuristic decay algorithm, taking de (Δ t) =1/Δ t when the time interval is short, taking de (Δ t) =1/log (e + Δ t) when the time interval is long,
9. The method according to claim 8, wherein in step S111, the multi-layered perceptron is a neural network combining a smart perceptron and a non-linear function, the neural network being:
the first loss function is:
wherein ,
is a hidden state set of similar users of the current user, is based on>
10. The method of claim 9, further comprising testing the recommended algorithm in step S113, wherein the testing environment of the testing comprises recording the testing results based on public data sets and on-line experiments, and comparing the results of the recommended algorithm with those of other models.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211533857.0A CN115953215B (en) | 2022-12-01 | 2022-12-01 | Search type recommendation method based on time and graph structure |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211533857.0A CN115953215B (en) | 2022-12-01 | 2022-12-01 | Search type recommendation method based on time and graph structure |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115953215A true CN115953215A (en) | 2023-04-11 |
| CN115953215B CN115953215B (en) | 2023-09-05 |
Family
ID=87286678
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211533857.0A Active CN115953215B (en) | 2022-12-01 | 2022-12-01 | Search type recommendation method based on time and graph structure |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115953215B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116522011A (en) * | 2023-05-16 | 2023-08-01 | 深圳九星互动科技有限公司 | Big data-based pushing method and pushing system |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN109522474A (en) * | 2018-10-19 | 2019-03-26 | 上海交通大学 | Recommended method based on interaction sequence data mining depth user's similitude |
| CN110516160A (en) * | 2019-08-30 | 2019-11-29 | 中国科学院自动化研究所 | User modeling method, the sequence of recommendation method of knowledge based map |
| CN111127142A (en) * | 2019-12-16 | 2020-05-08 | 东北大学秦皇岛分校 | Article recommendation method based on generalized neural attention |
| CN113190751A (en) * | 2021-05-10 | 2021-07-30 | 南京理工大学 | Recommendation algorithm for generating fused keywords |
| CN113722583A (en) * | 2021-07-31 | 2021-11-30 | 华为技术有限公司 | Recommendation method, recommendation model training method and related products |
| US20220207587A1 (en) * | 2020-12-30 | 2022-06-30 | Beijing Wodong Tianjun Information Technology Co., Ltd. | System and method for product recommendation based on multimodal fashion knowledge graph |
| CN114693397A (en) * | 2022-03-16 | 2022-07-01 | 电子科技大学 | Multi-view multi-modal commodity recommendation method based on attention neural network |
| CN115221387A (en) * | 2022-07-13 | 2022-10-21 | 全拓科技(杭州)股份有限公司 | Enterprise information integration method based on deep neural network |
-
2022
- 2022-12-01 CN CN202211533857.0A patent/CN115953215B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN109522474A (en) * | 2018-10-19 | 2019-03-26 | 上海交通大学 | Recommended method based on interaction sequence data mining depth user's similitude |
| CN110516160A (en) * | 2019-08-30 | 2019-11-29 | 中国科学院自动化研究所 | User modeling method, the sequence of recommendation method of knowledge based map |
| CN111127142A (en) * | 2019-12-16 | 2020-05-08 | 东北大学秦皇岛分校 | Article recommendation method based on generalized neural attention |
| US20220207587A1 (en) * | 2020-12-30 | 2022-06-30 | Beijing Wodong Tianjun Information Technology Co., Ltd. | System and method for product recommendation based on multimodal fashion knowledge graph |
| CN113190751A (en) * | 2021-05-10 | 2021-07-30 | 南京理工大学 | Recommendation algorithm for generating fused keywords |
| CN113722583A (en) * | 2021-07-31 | 2021-11-30 | 华为技术有限公司 | Recommendation method, recommendation model training method and related products |
| CN114693397A (en) * | 2022-03-16 | 2022-07-01 | 电子科技大学 | Multi-view multi-modal commodity recommendation method based on attention neural network |
| CN115221387A (en) * | 2022-07-13 | 2022-10-21 | 全拓科技(杭州)股份有限公司 | Enterprise information integration method based on deep neural network |
Non-Patent Citations (5)
| Title |
|---|
| LIN, JING 等: "FISSA: Fusing Item Similarity Models with Self-Attention Networks for Sequential Recommendation", 《RECSYS 2020: 14TH ACM CONFERENCE ON RECOMMENDER SYSTEMS》, pages 130 - 139 * |
| YANGYANG XU 等: "SSSER: Spatiotemporal Sequential and Social Embedding Rank for Successive Point-of-Interest Recommendation", 《IEEE ACCESS》, vol. 7, pages 156804 - 156823, XP011757021, DOI: 10.1109/ACCESS.2019.2950061 * |
| 潘冠源: "神经协同过滤模型的改进及其在推荐***中的应用", 《中国优秀硕士学位论文全文数据库 信息科技辑》, no. 1, pages 138 - 3336 * |
| 蒋忠元 等: "社交网络中的社团隐私研究综述", 《网络与信息安全学报》, vol. 7, no. 2, pages 10 - 21 * |
| 贾伟涛: "基于用户动态兴趣的推荐算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》, no. 5, pages 138 - 1508 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116522011A (en) * | 2023-05-16 | 2023-08-01 | 深圳九星互动科技有限公司 | Big data-based pushing method and pushing system |
| CN116522011B (en) * | 2023-05-16 | 2024-02-13 | 深圳九星互动科技有限公司 | Big data-based pushing method and pushing system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115953215B (en) | 2023-09-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wu et al. | Modeling the evolution of users’ preferences and social links in social networking services | |
| Lin et al. | A survey on reinforcement learning for recommender systems | |
| Chen et al. | Hybrid-order gated graph neural network for session-based recommendation | |
| CN113806630B (en) | Attention-based multi-view feature fusion cross-domain recommendation method and device | |
| Chulyadyo et al. | A personalized recommender system from probabilistic relational model and users’ preferences | |
| CN105760443A (en) | Project recommending system, device and method | |
| CN112632296B (en) | Knowledge graph-based paper recommendation method and system with interpretability and terminal | |
| CN114639483A (en) | Electronic medical record retrieval method and device based on graph neural network | |
| Gui et al. | Mention recommendation in twitter with cooperative multi-agent reinforcement learning | |
| CN115953215B (en) | Search type recommendation method based on time and graph structure | |
| Abugabah et al. | Dynamic graph attention-aware networks for session-based recommendation | |
| Liu et al. | Efficient hyperparameters optimization through model-based reinforcement learning and meta-learning | |
| Yang et al. | Intellitag: An intelligent cloud customer service system based on tag recommendation | |
| CN113449176A (en) | Recommendation method and device based on knowledge graph | |
| Yan et al. | Modeling long-and short-term user behaviors for sequential recommendation with deep neural networks | |
| CN113256024A (en) | User behavior prediction method fusing group behaviors | |
| He et al. | Using Cognitive Interest Graph and Knowledge-activated Attention for Learning Resource Recommendation | |
| Xu et al. | Similarmf: a social recommender system using an embedding method | |
| Jia et al. | A self-supervised learning framework for sequential recommendation | |
| Hou et al. | Hierarchical Transition-Aware Graph Attention Network for Session-based Recommendation | |
| Yan et al. | A short-term forecasting model with inhibiting normal distribution noise of sale series | |
| Anari et al. | Optimizing membership functions using learning automata for fuzzy association rule mining | |
| CN116521972B (en) | Information prediction method, device, electronic equipment and storage medium | |
| Lu et al. | Current Interest Enhanced Graph Neural Networks for Session-based Recommendation | |
| Hartatik et al. | A Comparison of BAT and firefly algorithm in neighborhood based collaborative filtering |
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 |