CN115953215B - 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
- CN115953215B CN115953215B CN202211533857.0A CN202211533857A CN115953215B CN 115953215 B CN115953215 B CN 115953215B CN 202211533857 A CN202211533857 A CN 202211533857A CN 115953215 B CN115953215 B CN 115953215B
- Authority
- CN
- China
- Prior art keywords
- user
- item
- search
- representing
- articles
- 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
- 238000000034 method Methods 0.000 title claims abstract description 51
- 239000013598 vector Substances 0.000 claims abstract description 92
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 72
- 230000006399 behavior Effects 0.000 claims abstract description 25
- 238000013528 artificial neural network Methods 0.000 claims abstract description 21
- 230000008447 perception Effects 0.000 claims abstract description 9
- 238000005070 sampling Methods 0.000 claims abstract description 6
- 238000013507 mapping Methods 0.000 claims abstract description 3
- 230000006870 function Effects 0.000 claims description 29
- 238000004364 calculation method Methods 0.000 claims description 20
- 238000002474 experimental method Methods 0.000 claims description 14
- 238000012360 testing method Methods 0.000 claims description 14
- 230000003044 adaptive effect Effects 0.000 claims description 13
- 238000005295 random walk Methods 0.000 claims description 13
- 230000007246 mechanism Effects 0.000 claims description 12
- 239000000047 product Substances 0.000 claims description 11
- 238000012549 training Methods 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 6
- 230000002708 enhancing effect Effects 0.000 claims description 4
- 239000013589 supplement Substances 0.000 claims description 4
- 230000008569 process Effects 0.000 abstract description 11
- 230000000694 effects Effects 0.000 description 9
- 238000012512 characterization method Methods 0.000 description 6
- 230000007774 longterm Effects 0.000 description 6
- 230000000052 comparative effect Effects 0.000 description 4
- 238000011156 evaluation Methods 0.000 description 4
- 238000012952 Resampling Methods 0.000 description 3
- 238000013459 approach Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 241000282326 Felis catus Species 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000001537 neural effect Effects 0.000 description 2
- 238000004091 panning Methods 0.000 description 2
- 238000012545 processing 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
- 230000003542 behavioural effect Effects 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
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000013135 deep learning 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
- 230000000737 periodic effect Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000002787 reinforcement Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000010561 standard procedure Methods 0.000 description 1
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 a time and graph structure, which relates to the field of recommendation systems and comprises the following steps: collecting user behavior historical data of an internet platform, encoding the user and the articles in the historical data by using an inner product neural network, calculating embedded vectors of the user 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 user 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 multi-layer perception machine to predict and back propagate, and applying the trained algorithm model to a recommendation algorithm. The recommendation method provided by the invention can process a large-scale user history sequence in an online environment, and can effectively improve 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 a time and graph structure.
Background
With the rapid increase in the length of user behavior history, how to effectively help users find items of interest from a vast number of candidate items on an internet platform has become a very important issue in the field of recommendation systems. Classical recommendation methods, such as collaborative filtering models and factorizer models, focus mainly on modeling the overall interests of a user in the hope of getting items of interest to the user. Such models rarely explore the needs of the user over a fixed period of time, which is one of the very important factors affecting the needs of the user. Therefore, there are some ways to attempt to capture the sequence pattern of the user through a memory network, a recurrent neural network or a time point model recently, but most of the methods cannot be applied in the actual internet environment due to the limitation of the computational complexity and the storage space when the history sequence of the user is too long.
Classical recommendation algorithms typically recommend items that a user wants through rich user-item interaction histories, typically in the form of tables, sequences, or graphs. However, according to the "click rate prediction user behavior search" published in the international information search institute of Qin Jiarui, as the user behavior data is increasingly accumulated, it is very difficult to train a model from the entire user log due to the limitation of online calculation. One possible approach is to focus only on the recent history of the user, using short-term history instead of long-term history, thereby generating personalized recommendations. But recent papers point out (e.g. "click through Rate prediction search based user interest modeling and long term sequential behavior" at information and knowledge management conferences) that these approaches fail to encode periodic and long term demand dependencies for users, so focusing only on the recent history of users is actually a suboptimal solution. The two papers mentioned above then suggest that hard and soft searches can be performed throughout the user history. Qin Jiarui also proposes modeling query construction for context data by reinforcement learning, and then using this module to query search related behavior using BM25 relevance functions. Papers have also been proposed (search-based user interest modeling and long-term sequential behavior for click rate prediction) to match related items through hard retrieval of item categories and to soft retrieve matching related items based on item characterization vectors. In this way, the model can make recommendations using relevant items retrieved throughout the user's behavioral history. Existing search-based methods ignore time intervals in the user behavior history. And the related items that are typically retrieved are all positive feedback (e.g., the user clicks on the item), which gives the opportunity to consider both positive and negative feedback related items to make full use of the entire user sequence.
Articles such as "modeling user behavior with time LSTM model" and Hou Saini published on the Manual and data engineering journal "time recommended recursive Poisson factorization" by the aid of time intervals between user behaviors, which is not available with conventional sequence structures, are published on the International Association of artificial intelligence. One direction is to control short-term and long-term interest updates by using new dedicated time gates, such as Zhao Pengpeng, a space-time LSTM model for recommending next interest points, proposes a cyclic neural network-based distance gate to control short-term and long-term interest point updates. Another way to use time interval information is to specify the user's sequential history through a point process, i.e., discrete times in the user's history are modeled as continuous times, such as Mei Hongyuan et al published in the journal of the journal information processing systems, the process of neural houx: a neural self-regulating multivariable point process suggests a hough process that allows past events to influence future predictions through a complex and realistic style. In addition to the high computational complexity and high 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 impractical 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, and the method has great application in a deep learning method. The principal feature learning of the main flow is to build more generalized node adjacency relations, either by modeling the adjacency matrix of the graph or by randomly walk and fit the actual adjacency relations. The most common deep walk model is one of the models that uses random walk at the earliest, by randomly walking in a directed graph to obtain a sufficient number of node sequences, and considering these node sequences as a sentence similar to natural language, inputting into a jump graph model to learn the embedded vector of the node. The node vector model is a variant of the deep walk model, which controls the tendency of the random walk to become a biased random walk by two parameters, p and q. In addition, the learning model is characterized by an additional information enhancement graph, the node graph of the article is built by continuous article access behaviors of hundreds of millions of users on the Internet, the node sequence is sampled by random walk, and the embedded vector of the node is trained by a jump graph model with characteristics. Wang Hongwei in the journal of knowledge and data engineering, "learning graph characterization with generated challenge network," proposes a model for graph characterization learning through a challenge generation network, which first performs policy random walk and restart on the graph based on a learnable graph operator, samples node pairs of a center node and a last node before restart, and then learns by a arbiter to determine whether a pair of nodes has edges. The embedded vector of the node is learned by this countermeasure method.
For the summary of related researches at home and abroad, the following steps are: at present, the problem that the online calculation is too complicated when the user behavior history is too long is difficult to solve by using a wide recommendation algorithm in the industry, meanwhile, the common algorithm cannot effectively utilize the user behavior history, and the problems of neglecting the time interval of the user behavior, negative feedback of the user and the like exist. Further, in recent time-aware algorithms, it is often difficult to solve the problem of large noise caused by direct input of time intervals.
Accordingly, those skilled in the art have been working to develop a model with more efficient use of the user's historical behavior and more reasonable time perception, which performs better than traditional recommendation algorithms in more situations.
Disclosure of Invention
In view of the above-mentioned drawbacks of the prior art, the present invention is to provide a search-based prediction model, which fuses feedback information and historical sequence information of an object in historical data of a user, so as to improve the performance of the model on a long-tail user.
In order to achieve the above object, the present invention provides a search recommendation method based on a time and graph structure, which is characterized in that the method comprises the following steps:
s101: collecting user behavior historical data of an internet platform, encoding the user and 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 the 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 characteristics;
s105: retrieving similar items of a target item in the user history data using a search model based on the user and the item characteristics;
s107: obtaining similar users of the target user by using the search model, and retrieving similar objects of the target object 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 circulation 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 multi-layer perceptron for prediction, and carrying out back propagation according to a first loss function;
s113: and applying the trained algorithm model to a recommendation algorithm.
Further, in the step S101, when the inner product neural network is used to calculate the embedded vector, the following manner is adopted:
wherein ,an embedded vector representing the nth item, +.>Attribute vector representing the nth item, +.> and />Representing weights during calculation +.>The p-th item in the attribute vector representing the n-th item.
Further, in the step S103, when the embedded vector of the item is obtained, an expansion algorithm for enhancing the graph feature learning is adopted, and an additional information enhancement graph embedded model is used to model the item sequence in the user history.
Further, when the historical articles of the user are mapped in step S103, positive samples and negative samples are sampled in the map, the positive samples are resampled through random walk, the types of the articles are 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 that the embedding vectors of the articles are obtained.
Further, the random walk is a walk-in-node learning approach, trained using the second loss function after the positive and negative samples are sampled, and the item-embedding vector is calculated using the following formula:
the second loss function is:
wherein ,an embedded vector representing the nth item, +.>Is a weight parameter, ++>Representing i m Negative examples of articles.
Further, in step S105, the search model uses an adaptive search algorithm that supports a hard search, which is a search in which the search item types are identical, and a soft search, which is a search in which the search item types are not similar at all, and calculates the similarity between the target item type and the user history item type by:
wherein ,representing a collection of items of a similar type to the target item in the history of the mth user, H m Represents the history of the mth user, +.>Representing recent history, i, in the mth user n Identification number representing the nth item, +.>Indicates the type of the nth object item, +.>Indicating the type of the search object of the nth target object, x i An attribute vector representing an item, wherein->A p-th item in an attribute vector representing an n-th item; τ is used to control the degree of similarity, the greater τ, the closer τ is to similarity, and τ is reduced from its magnitude during training.
Further, in step S107, the similarity between users is calculated by calculating the same kind of articles in the histories of the target user and other users, so as to retrieve the similar users of the target user, the historic behaviors of the similar users are used as additional information to supplement, and the similarity between users is calculated by adopting the following calculation method:
wherein ,represents a set of users similar to the mth user, U represents the entire users, H m Represents the history of the mth user, u m An identification number representing the mth user, +.>Representing the class of the mth target user, +.>Representing the category of the search user of the mth target user,/->Representing the embedding vector of the mth user, iota is used to control the degree of similarity, the greater iota, the closer to similarity, and the smaller iota is from large during training.
Further, in step S109, a preliminary result of the hidden state is obtained by calculation by the gate cycle unit, the preliminary result is further calculated by the attention mechanism to obtain the hidden state, and the hidden state set is aggregated by an aggregation function;
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 ,representing the t-th item feature vector, +.>A user feedback vector representing the user for the t-th item,is-> and />Vector formed by splicing, h' t and h′t-1 Is a preliminary hidden state, h t Is in a hidden state, W x and Ux As a weight parameter Δt is two items i t-1 and it The time interval between the two is a heuristic attenuation algorithm, the time of taking de (delta t) =1/delta t when the time interval is short, the time of taking de (delta t) =1/log (e+delta t) when the time interval is long,
representing a set of hidden states.
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 ,a set of hidden states for similar users to the current user, < ->
Further, in step S113, the method further includes testing the recommendation algorithm, where a testing environment of the test includes recording the test 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 present invention, compared with the prior art, the present invention has the following beneficial effects:
1. the sequence recommendation algorithm based on search and time perception, which is provided by the invention, gradually transits from hard search to soft search, so that the search process is more reasonable, and compared with the traditional classical sequence recommendation algorithm (comprising the time-perception sequence recommendation algorithm), the sequence recommendation algorithm can process a large-scale user history sequence in an online environment, and the recommendation efficiency and performance can be effectively improved;
2. according to the invention, the historical feedback of the user is used as additional information to be input into the model, so that the model has a better effect, the user-article is modeled by a graph representation learning method, the association between articles is captured by learning the embedded vector of the article, and the method has a better effect on three classical data sets from real Internet application compared with the advanced recommendation algorithm at home and abroad.
The conception, specific structure, and technical effects of the present invention will be further described with reference to the accompanying drawings to fully understand the objects, features, and effects of the present invention.
Drawings
FIG. 1 is a flow chart of a preferred embodiment of the present invention;
FIG. 2 is a comparison experiment result on an offline public data set according to a preferred embodiment of the present invention;
FIG. 3 is a comparative experiment result on an online scene of a preferred embodiment of the present invention;
FIG. 4 is a comparison of experimental results of the algorithm herein for two configurations on an online scenario in accordance with a preferred embodiment of the present invention;
FIG. 5 is a graph showing the results of a comparative experiment of the expansion algorithm and the original algorithm in the online scene according to a preferred embodiment of the present invention.
Detailed Description
The following description of the preferred embodiments of the present invention refers to the accompanying drawings, which make the technical contents thereof more clear and easy to understand. The present invention may be embodied in many different forms of embodiments and the scope of the present invention is not limited to only the embodiments described herein.
In the drawings, like structural elements are referred to by like reference numerals and components having similar structure or function are referred to by like reference numerals. The dimensions and thickness of each component shown in the drawings are arbitrarily shown, and the present invention is not limited to the dimensions and thickness of each component. The thickness of the components is exaggerated in some places in the drawings for clarity of illustration.
As shown in fig. 1, the search recommendation method based on time and graph structure provided by the embodiment of the invention includes the following steps:
s101: and collecting user behavior historical data of the Internet platform, encoding the user and the articles in the historical data by using the 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 to calculate the embedded vector, the following manner is adopted:
wherein ,an embedded vector representing the nth item, +.>Attribute vector representing the nth item, +.> and />Representing weights during calculation +.>The p-th item in the attribute vector representing the n-th item.
S103: and (3) mapping, sampling and learning historical articles of the user, capturing the associated information among the articles, and inputting the embedded vectors of the articles into the model as article characteristics.
When the embedded vector of the object is obtained, an expansion algorithm for enhancing the characteristic learning is adopted, and an additional information enhancement graph embedded model is used for modeling the object sequence in the user history. When a diagram is built on a historical object of a user, positive samples and negative samples are sampled in the diagram, the positive samples are resampled through random walk, the object types are used as additional information, the positive samples, the negative samples and the additional information are input into an additional information enhancement diagram embedding model for learning, and an embedding vector of the object is obtained. When resampling the positive sample by a random walk, the random walk is a walk mode in node embedding learning, after sampling the positive sample and the negative sample, training is performed by using the second loss function, 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 that can be learned, +.>Representing i m Negative examples of articles.
S105: based on the user and item characteristics, similar items of the target item are retrieved in the user history data using a search model.
An item similar to the target item is retrieved in the user history by calculating a similarity between the category of the target item and the category of the user history item. The calculation of the item similarity is adaptive, and the higher-demand hard similarity (in the extreme case, the item types are required to be identical) gradually transits to the soft similarity (in the extreme case, the similarity is not identical) along with the training process. The above-mentioned search model uses an adaptive search algorithm supporting a hard search, which is a search in which the types of search items are identical, and a soft search, which is a search in which the types of search items are completely dissimilar, the adaptive search algorithm calculating the similarity between the types of the target item and the types of the user's history item in the following manner:
wherein ,representing a collection of items of a similar type to the target item in the history of the mth user, H m Represents the history of the mth user, +.>Representing recent history, i, in the mth user n Identification number representing the nth item, +.>Indicates the type of the nth object item, +.>Indicating the type of the search object of the nth target object, x i An attribute vector representing an item, wherein->A p-th item in an attribute vector representing an n-th item; τ is used to control the degree of similarity, the greater τ, the closer τ is to similarity, and τ is reduced from its magnitude during training.
S107: and obtaining similar users of the target user by using the search model, and retrieving similar objects of the target object from historical data of the similar users.
The 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 the similar users of the target user are searched, and the historic behaviors of the similar users are used as additional information to supplement.
The similarity between users adopts the following calculation mode:
wherein ,represents a set of users similar to the mth user, U represents the entire users, H m Represents the history of the mth user, u m An identification number representing the mth user, +.>Representing the class of the mth target user, +.>Representing the category of the search user of the mth target user,/->Representing the embedding vector of the mth user, iota is used to control the degree of similarity, the greater iota, the closer to similarity, and the smaller iota is from large during training.
S109: and inputting the retrieved user history sequence into a time perception model, and obtaining a hidden state sequence after passing through a gate circulation unit and an attention mechanism, wherein the user history sequence comprises information such as user history feedback information, user history time interval and the like.
And calculating to obtain a preliminary result of the hidden state through the gate circulation unit, further calculating to obtain a final hidden state through an attention mechanism by the preliminary result, and aggregating the hidden state set through an aggregation function.
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 ,representing the t-th item feature vector, +.>A user feedback vector representing the user for the t-th item,is-> and />Vector formed by splicing, h' t and h′t-1 Is a preliminary hidden state, h t Is in a hidden state, W x and Ux As a weight parameter Δt is two items i t-1 and it The time interval between the two is a heuristic attenuation algorithm, the time of taking de (delta t) =1/delta t when the time interval is short, the time of taking de (delta t) =1/log (e+delta t) when the time interval is long,
representing a set of hidden states.
S111: the hidden state sequence is input into a multi-layer perceptron for prediction and back-propagation is performed according to a first loss function.
The multi-layer 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 ,a set of hidden states for similar users to the current user, < ->
S113: and applying the trained algorithm model to a recommendation algorithm.
And testing the recommendation algorithm, wherein the test comprises the steps of recording test results based on a public data set and an online experiment, and comparing the result difference between the recommendation method and other models.
The search type recommendation method based on the time and 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, which can process a large-scale user history sequence in an online environment compared with the traditional classical sequence recommendation algorithm (comprising the time perception sequence recommendation algorithm).
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. The invention inputs the historical feedback of the user into the model as additional information, and the model using the skill has better experimental effect.
4. In the invention, a method of feature learning is used to model a user-item, and the association between items is captured by learning the embedded vector of the item.
5. The method can obtain good effect compared with the advanced recommendation algorithm at home and abroad on three classical data sets from real Internet application, and can effectively improve recommendation efficiency and performance in online recommendation scenes.
As shown in fig. 2 to fig. 5, for the search type recommendation method based on time and graph structure provided by the embodiment of the present invention, in order to verify the experimental result of the present invention, the present invention shows comparative experimental results in three types of real offline public data sets and one real online recommendation scene, and for the offline public data sets, the comparative baseline algorithm is 3 standard methods commonly used in recommendation algorithms: a deep interest network, a deep interest evolution network and a search-based behavior modeling recommendation algorithm.
The method is characterized in that a comparison test is carried out on three public data sets of a cat, a payment bank and a panning bank and a sponsor bank 'guessing you like' recommended scene respectively, wherein the first three data sets come from the application of a real Internet platform, and the 'guessing you like' recommended scene is an online scene which is very important in the industry. In an offline dataset comparison experiment, main evaluation indexes are area under ROC curve (AUC), accuracy (ACC) and log loss value (LogLoss), "text algorithm-" is a configuration without using label skills in the algorithm of the invention, and "text algorithm+" is a configuration using graph characterization learning in the algorithm of the invention; for a 'you like' online recommendation scene, main evaluation indexes are Click Through Rate (CTR), AUC, average article 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 in subscripts are results obtained after users without clicking are removed, meanwhile, barrel statistics is carried out on expansion experiments, comparison of experimental results of users with different history lengths under two algorithms is carried out (the difference value of CTR refers to the CTR value of a CTR result subtraction algorithm of the expansion algorithm), and the overall comparison test effect is shown in figures 1,2,3 and 4. From the graph, the result of the invention compared with the baseline recommendation algorithm can obtain better results on benefit indexes, which indicates that the technical scheme of the invention is more effective, and the expansion algorithm has better results compared with the original algorithm.
The present invention will be described in detail with reference to preferred embodiments thereof.
As shown in fig. 1, the implementation flow of the present invention includes the following steps:
and firstly, collecting real user behavior histories in an Internet platform, encoding types of a user and an article by using an inner product neural network, and uniformly processing.
In this step, the basic elements are defined as follows:
(1) user identification number u and item identification number i: a vector representing unique identification numbers for each different user and item, where u m Representing the identification number of the mth user, i n An identification number representing an nth item;
(2) user attribute vector x u And an item attribute vector x i : representing attribute vectors of users and items, whereinRepresenting the p-th item in the attribute vector of the mth user, and +.>A p-th item in an attribute vector representing an n-th item;
(3) user embedded vector e u And an item embedding vector e i : representing embedded vectors of a user and an item, whereinAn embedded vector representing the mth user, and +.>An embedded vector representing an nth item;
the inner product neural network is used on this basis to calculate the embedded vector:
wherein and />Representing the weight that can be learned during the calculation, < +.>The calculation method is the same as that of the previous step.
And secondly, using an expansion algorithm for enhancing the characteristic learning to build, sample and learn historical articles of the user, capturing the associated information among the articles, and taking the learned article embedded vectors as article characteristics to be input into a model together.
The additional information enhancement map embedding model is used to model the sequence of items in the user history, obtaining the embedding vector of its items. The method comprises the steps of building a graph of objects in a user history, sampling positive and negative samples in the graph, resampling positive samples through random walk, taking the types of the objects as additional information, learning the obtained positive samples, negative samples and additional information through an additional information enhancement graph embedding model, obtaining embedded vectors, and processing the embedded vectors together as additional features of the objects. The walk mode used for resampling is a walk mode in node embedding learning, and the mode for calculating the article embedding vector is as follows:
wherein ,an embedded vector representing the nth item, +.>Is a weight parameter which can be learned, the second lossThe loss function is:
wherein ,representing i m Negative examples of articles.
And thirdly, searching the object with the object with high similarity in the history of the user by using an adaptive algorithm based on the user and the object characteristics obtained in the first and second steps by using a search-based model for the object user and the object.
The user and item code pairs obtained in the first and second steps are used for searching the items in the user history which are needed to be similar to the target items through an adaptive search-based module. An item similar to the target item is retrieved in the user history by calculating a similarity between the category of the target item and the category of the user history item. The calculation of the item similarity is adaptive, and the higher-demand hard similarity (in the extreme case, the item types are required to be identical) gradually transits to the soft similarity (in the extreme case, the similarity is not identical) along with the training process. Under such an adaptive algorithm, the overall algorithm architecture will form an end-to-end style. In addition to calculating the similarity of the target item type and the user's historical item type, this adaptive algorithm will also be applied to find similar users. The 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 the users similar to the target user are searched, and the historic behaviors of the users are used as additional information to supplement the additional information, so that a better effect is obtained.
On this basis, the basic elements are defined as follows:
(1) item attribute vector x i : representing attribute vectors of an item, whereinRepresenting the attribute vector of the nth articleItem p;
(2) user embedded vector e u : representing an embedded vector of a user, whereinAn embedded vector representing an mth user;
(3) user class c u And article category c i : representing the category of users and articles, whereinRepresenting the class of the mth target user, +.>Representing the category of the search user of the mth target user,/->Indicates the type of the nth object item, +.>A search item type indicating an nth target item;
(4) user identification number u and item identification number i: a vector representing unique identification numbers for each different user and item, where u m Representing the identification number of the mth user, i n An identification number representing an nth item;
on this basis, the items requiring similarity are retrieved:
wherein ,calendar representing mth userCollections of items of similar kind to the target item in the history, H m Represents the history of the mth user, +.>Representing recent history in the mth user, τ is used to control the degree of hard and soft similarity, the greater τ, the closer τ is to the hard similarity, and τ is reduced from large during training.
The method of finding similar users is similar:
wherein ,representing a set of users similar to the mth user, U representing the totality of users, iota being used to control the degree of hard and soft similarity, the greater iota, the closer the hard similarity is, and the smaller iota is from large to small during training.
And step four, similar to the step three, obtaining users with high similarity with the target users by using an adaptive algorithm, and retrieving articles with high similarity with the target articles in the histories of the users.
For a given user u m And a given target object i n Thereafter, a set of similar items can be retrieved by step 3And similar user->Can write into->For->Each item i of (1) n We know user u m Feedback (click or not, etc.) on the item. The feature vector of the item and the user feedback click are then input into a gate cycle unit for learning to discover useful sequence patterns. After passing through the gate cycle unit, the change in user interest is modeled using an attention mechanism, and this feature of time interval is taken into account. The hidden state sequence obtained after passing through the gate cycle unit and the attention mechanism is obtained by a designed aggregation function to obtain a vector for representing the hidden state sequence. On this basis, the following basic elements are defined:
(1) article feature vectorA vector representing a characteristic of the t-th item;
(2) user feedback vectorRepresenting an embedded vector obtained according to feedback of a user on the t-th article;
inside the door cycle unit we calculate 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 and,is-> and />Spliced vectors, 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 weight parameter which can be learned, h t Is in a hidden state. Δt is two items i t-1 and it The time interval between the two is a heuristic attenuation algorithm, wherein the time of the time interval is shorter, the time of the time interval is de (delta t) =1/delta t, and the time of the time interval is longer, the time of the time interval is de (delta t) =1/log (e+delta t).
For each setClosing deviceWe 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 +.> wherein />
And fifthly, inputting the retrieved user history sequence into a time-aware model, wherein the user history feedback information and the user history time interval are used as features to be input into the model.
The final neural network 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 multi-layer perceptron to predict, and carrying out back propagation according to the first loss function.
And step seven, applying the trained algorithm model to a public data set and an online experiment, recording the experimental result of the algorithm model, and comparing the difference between the experimental result and the results of other models.
According to the search type recommendation method based on the time and graph structure, comparison experiments are conducted in three types of real offline public data sets and a real online recommendation scene, and the test results are shown in fig. 2-5. In a comparison experiment, the offline public data set selects three applications from a real Internet platform, such as a cat, a payment treasures and a panning treasures, and the online scene selects a "guessing you like" recommended scene from a sponsored bank (the "guessing you like" recommended scene is an online scene which is very important in the industry). For the offline public dataset, the baseline algorithm for comparison is the 3 baseline methods commonly used in the recommendation algorithm: a deep interest network, a deep interest evolution network and a search-based behavior modeling recommendation algorithm. In an offline dataset comparison experiment, main evaluation indexes are area under ROC curve (AUC), accuracy (ACC) and log loss value (LogLoss), "text algorithm-" is a configuration without using label skills in the algorithm of the invention, and "text algorithm+" is a configuration using graph characterization learning in the algorithm of the invention; for a 'guessing you like' online recommendation scene, main evaluation indexes are Click Through Rate (CTR), AUC, average article 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 as subscripts are obtained after users without clicking are removed, meanwhile, barrel statistics is carried out on expansion experiments, and comparison of experimental results of users with different history lengths under two algorithms is carried out (the difference value of CTR refers to the CTR value of a CTR result subtraction algorithm of the expansion algorithm). From fig. 2-5, it can be observed that the results of the present invention compared with the baseline recommendation algorithm can obtain better results on the benefit index, which indicates that the technical scheme of the present invention is more effective, and the expansion algorithm has better results compared with the original algorithm.
The foregoing describes in detail preferred embodiments of the present invention. It should be understood that numerous modifications and variations can be made in accordance with the concepts of the invention without requiring creative effort by one of ordinary skill in the art. Therefore, all technical solutions which can be obtained by logic analysis, reasoning or limited experiments based on the prior art by the person skilled in the art according to the inventive concept shall be within the scope of protection defined by the claims.
Claims (8)
1. A search recommendation method based on a time and graph structure, the method comprising the steps of:
s101: collecting user behavior historical data of an internet platform, encoding the user and 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 the 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 characteristics;
s105: retrieving similar items of a target item in the user history data using a search model based on the user and the item characteristics;
s107: obtaining similar users of the target user by using the search model, and retrieving similar objects of the target object 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 circulation 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 multi-layer perceptron for prediction, and carrying out back propagation according to a first loss function;
s113: applying the trained algorithm model to a recommendation algorithm;
wherein ,
the search model uses an adaptive search algorithm, the adaptive search algorithm supports hard search and soft search, the hard search is search with completely identical search item types, the soft search is search with completely dissimilar search item types, and the adaptive search algorithm calculates the similarity between a target item type and the user historical item type in the following manner:
wherein i is an article identification number,representing a set of items of a similar type to the target item in the history of the mth user, i n′ Identification number of search item for nth target item, H m Represents the history of the mth user, +.>Representing recent history, i, in the mth user n Identification number representing the nth item, +.>Indicates the type of the nth object item, +.>Indicating the type of the search object of the nth target object, x i Attribute vector representing an item->The p-th item in the attribute vector representing the n-th item, τ is used to control the degree of similarity, and the greater τ, the closer τ is to the similarity, and τ is reduced from large during training；
In step S107, the similarity between users is calculated by calculating the same kind of articles in the histories of the target user and other users, so as to retrieve the similar users of the target user, the historic behaviors of the similar users are used as additional information to supplement, and the similarity between users is calculated by adopting the following calculation method:
wherein u is the user identification number,represents a set of users similar to the mth user, U represents the entire users, H m Represents the history of the mth user, u m An identification number representing the mth user, +.>Representing the class of the mth target user, +.>Representing the category of the search user of the mth target user,/->Representing the embedding vector of the mth user, iota is used to control the degree of similarity, the greater iota, the closer to similarity, and the smaller iota is from large during training.
2. The method according to claim 1, wherein in the step S101, when calculating the embedded vector using the inner product neural network, the following manner is adopted:
wherein ,an embedded vector representing the nth item, +.>Attribute vector representing the nth item, +.> and />Representing weights during calculation +.>The p-th item in the attribute vector representing the n-th item.
3. The method according to claim 2, wherein in step S103, an expansion algorithm for enhancing graph feature learning is used to model the sequence of items in the user history using an additional information enhancement graph embedding model when acquiring the embedding vector of the items.
4. The method of claim 3, wherein when mapping the historical items of the user in step S103, positive samples and negative samples are sampled in the map, the positive samples are resampled by random walk, the item types are 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 item.
5. The method of claim 4, wherein the random walk is a walk in node embedding learning, trained using a 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, ++>Representing i m Negative sample of the item, i is the item identification number, i n Indicating the identification number of the nth article.
6. The method according to claim 5, wherein in step S109, a preliminary result of the hidden state is obtained by calculation by the gate cycle unit, the preliminary result is further calculated by the attention mechanism to obtain the hidden state, and the hidden state set is aggregated by an aggregation function;
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 ,representing the t-th item feature vector, +.>User feedback vector representing user to item t,>is that and />Vector formed by splicing, h' t and h′t-1 Is a preliminary hidden state, h t Is in a hidden state, W x and Uxy As a weight parameter Δt is two items i t-1 and it The time interval between the two is a heuristic attenuation algorithm, the time of taking de (delta t) =1/delta t when the time interval is short, the time of taking de (delta t) =1/log (e+delta t) when the time interval is long,
representing a set of hidden states.
7. The method of claim 6, wherein in step S111, the multi-layer perceptron is a neural network of intelligent perceptrons and nonlinear functions, the neural network being:
the first loss function is:
wherein ,a set of hidden states for similar users to the current user,
8. the method of claim 7, further comprising testing the recommended algorithm in step S113, the testing environment of the test comprising recording the test results based on a public dataset and on-line experiments, and comparing the recommended method with the differences in results 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 CN115953215A (en) | 2023-04-11 |
| CN115953215B true 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) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
Citations (8)
| 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 |
| 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 |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11935106B2 (en) * | 2020-12-30 | 2024-03-19 | Beijing Wodong Tianjun Information Technology Co., Ltd. | System and method for product recommendation based on multimodal fashion knowledge graph |
-
2022
- 2022-12-01 CN CN202211533857.0A patent/CN115953215B/en active Active
Patent Citations (8)
| 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 |
| 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 (1)
| Title |
|---|
| 神经协同过滤模型的改进及其在推荐系统中的应用;潘冠源;《中国优秀硕士学位论文全文数据库 信息科技辑》(第1期);第I138-3336页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115953215A (en) | 2023-04-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Liu et al. | Multi-behavioral sequential prediction with recurrent log-bilinear model | |
| Chen et al. | Deep reinforcement learning in recommender systems: A survey and new perspectives | |
| Lin et al. | A survey on reinforcement learning for recommender systems | |
| Hofmann | Fast and reliable online learning to rank for information retrieval | |
| Deng et al. | G^ 3SR: global graph guided Session-based recommendation | |
| Zhang et al. | GACOforRec: Session-based graph convolutional neural networks recommendation model | |
| CN115953215B (en) | Search type recommendation method based on time and graph structure | |
| Wang et al. | Incorporating link prediction into multi-relational item graph modeling for session-based recommendation | |
| CN114168819B (en) | Post matching method and device based on graph neural network | |
| Abugabah et al. | Dynamic graph attention-aware networks for session-based recommendation | |
| Yang et al. | Intellitag: An intelligent cloud customer service system based on tag recommendation | |
| CN113987343A (en) | Diversified recommendation method, system and storage medium based on reinforcement learning | |
| Xu et al. | Collective vertex classification using recursive neural network | |
| CN116719947A (en) | Knowledge processing method and device for detecting power inspection defects | |
| Han | Reinforcement learning approaches in dynamic environments | |
| Kancharla et al. | Applicability of automl to modeling of time-series data | |
| Qiao et al. | Sequence and graph structure co-awareness via gating mechanism and self-attention for session-based recommendation | |
| CN114841765A (en) | Sequence recommendation method based on meta-path neighborhood target generalization | |
| Ghosh et al. | DiPS: Differentiable Policy for Sketching in Recommender Systems | |
| CN115293812A (en) | E-commerce platform session perception recommendation prediction method based on long-term and short-term interests | |
| Chen et al. | Attentive capsule graph neural networks for session-based recommendation | |
| He et al. | Using Cognitive Interest Graph and Knowledge-activated Attention for Learning Resource Recommendation | |
| Anumasa et al. | Continuous Depth Recurrent Neural Differential Equations | |
| Menon et al. | Discrete Sequencing for Demand Forecasting: A novel data sampling technique for time series forecasting | |
| Hema et al. | Novel optimised deep learning approach for an efficient traffic state prediction based on CAE-ICCDCS-GRU model |
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 |