Graph convolution network session recommendation method and system with dwell time

Abstract

The application provides a graph convolution network conversation recommendation method and system with a residence time, obtains items involved in a conversation and a historical conversation sequence with a residence time, scales the residence time in the historical conversation sequence, constructs a directed conversation graph, sets the weight of an edge in the conversation graph according to the residence time and the number of occurrences, extracts a feature vector in the conversation graph through a graph convolution network, repeatedly screens features by using a feature selection method based on a penalty term, captures long-term and short-term preferences of a user by using an attention mechanism, and predicts the preferences of the user for items. The application combines the residence time into the conversation graph as an input of the graph convolution network, and the obtained node vector features are input into a target attention network as an input of a feature selection layer, so that overfitting is prevented and the accuracy of recommendation is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of computer technology recommendation algorithm, and particularly relates to a graph convolution network session recommendation algorithm with a residence time. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] With the development of the Internet, various kinds of information emerge in an endless stream. Information overload makes it extremely difficult for people to find the information they need, and the emergence of a recommendation system solves this problem. Session-based recommendation algorithms help users match different needs and have been widely used in streaming media, e-commerce and other fields, and have a high practical value.

[0004] In real life, short video, Taobao shopping and home appliance industry chain scenarios are recommended according to historical records; the residence time of a user on a project reflects the degree of love for the project, and the longer the residence time, the higher the degree of love, such as shopping websites like Jingdong and Taobao, which usually record the browsing time of users, and analyze the preferences of users according to the browsing time of users. Compared with recommending users according to the number of clicks, the efficiency of recommendation is improved. Traditional recommendation algorithms recommend according to the historical records of users, ignoring the potential relationship between projects, and session-based recommendation is based on the session sequence of users within a period of time, which most likely reflects the short-term interest preferences of users. Session recommendation based on graph neural networks not only reflects the short-term interest preferences of users, but also mines the potential relationship between projects, predicts the long-term preferences of users, and accurately recommends by combining the short-term preferences and long-term preferences of users.

[0005] Wu, Zhu et al. proposed a graph neural network recommendation algorithm based on sessions, applied graph neural networks to session-based recommendation, and proved the superiority of graph neural networks in recommendation; Wang, Cai et al. proposed a session-based attention-enhanced graph neural network recommendation algorithm, which uses target attention networks to capture specific user interests related to target projects and uses self-attention mechanisms to capture different priorities of different projects; but these two recommendation algorithms have the following problems: they only fixedly construct session sequences into session graphs, ignoring the possibility of introducing some features into the session graph, resulting in poor recommendation effect, and the node vector features obtained through the graph neural network are prone to overfitting problems. SUMMARY

[0006] In order to overcome the above-mentioned deficiencies of the prior art, the application provides a graph convolution network session recommendation method with a residence time, which combines the residence time into a session graph as an input of a graph convolution network, and obtains a node vector feature as an input of a target attention network through a feature selection layer, so as to prevent overfitting and improve the accuracy of recommendation.

[0007] To achieve the above object, one or more embodiments of the application provide the following technical solutions:

[0008] The application provides a graph convolution network session recommendation method with a residence time in a first aspect;

[0009] The graph convolution network session recommendation method with a residence time comprises:

[0010] Obtaining items and a session sequence, and scaling residence times in the session sequence;

[0011] Based on the session sequence scaled based on the residence times, a directed session graph is constructed, and the weights of edges in the session graph are set according to the residence times and the occurrence times of the edges;

[0012] Features in the session graph are extracted through a graph convolution network, and repeated features are screened through a feature selection method based on a penalty term;

[0013] User long-term and short-term preferences are captured through an attention mechanism, and items are recommended according to the user long-term and short-term preferences.

[0014] Further, the user interaction sequence S = {v s,1 ,v s,2 ,......v s,n} is a set of items sorted according to timestamps, the residence times T = {t1,t2,......,t n} represent the residence time distribution of the session sequence S, and t n ∈T represents the residence time from the item v s,n to the item v s,n+1 .

[0015] Further, the residence times are scaled through two methods of normalization and reverse normalization.

[0016] Further, a directed session graph G S =(V S ,E S ) is constructed, v s,i ∈V represents an item, each edge (v s,i-1 ,v s,i ) represents that the user clicks v s,i-1 after clicking v s,i in a session, and a standardized weight is set for each edge.

[0017] The calculation manner of the weight is that the residence time is multiplied by the occurrence times, and the strength of the relationship between two nodes is represented.

[0018] Further, the extraction manner of the node feature vector in the graph convolution network comprises the following processes:

[0019] According to the session graph, the item feature vector is extracted;

[0020] Based on the item feature vector, the neighbor information of the node is aggregated through an activation function;

[0021] The reserved feature information is obtained by using a forget gate, an input gate and an output gate;

[0022] The candidate information in the reserved feature information is obtained by using an activation function;

[0023] The information reserved by the forget gate and the candidate information selected by the input gate are merged to obtain final new information;

[0024] The output gate is used to select and output the new information.

[0025] Further, the feature selection is based on a logistic regression model with an L1 penalty term, and the features with the same recommendation value are reserved, and the logistic regression model with an L2 penalty term is used for optimization.

[0026] Further, the feature selection specific logic is that if the feature weight is 1 in L1, a similar set is built, if the weight difference of the features in the set is not large in L2 and the weight in L1 is 0, the features in the set are divided into equal weights in L1.

[0027] The second aspect of the application provides a graph convolution network session recommendation system with residence time.

[0028] The graph convolution network session recommendation system with residence time comprises:

[0029] The residence time scaling module is configured to obtain items and session sequences, and scale the residence time in the session sequences;

[0030] The session graph construction module is configured to construct a directed session graph based on the session sequences after residence time scaling, and set the weight of the edge in the session graph according to the residence time and the occurrence times of the edge;

[0031] The feature extraction and selection module is configured to extract the features in the session graph by the graph convolution network, and screen the repeated features by the feature selection method based on the penalty term;

[0032] The item recommendation module is configured to capture long-term and short-term preferences of the user by using an attention mechanism, and to make item recommendations according to the long-term and short-term preferences of the user.

[0033] The third aspect of the present application provides a computer readable storage medium, having a program stored thereon, which, when executed by a processor, implements the steps in the method for session recommendation of a graph convolutional network with dwell time according to the first aspect of the present application.

[0034] The fourth aspect of the present application provides an electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor implements the steps in the method for session recommendation of a graph convolutional network with dwell time according to the first aspect of the present application when executing the program.

[0035] The above one or more technical solutions have the following beneficial effects:

[0036] The present application introduces dwell time into the graph neural network session recommendation, and adds a feature selection layer, thereby improving the accuracy and recall rate of recommendation.

[0037] The representation of the session sequence preprocessing is improved, and the dwell time is combined into the session graph, which not only reflects the user's interest in the product, but also reflects the potential relationship between items.

[0038] The feature selection layer is introduced to solve the overfitting problem, and the target attention network is used to capture the specific user interest related to the target item, the self-attention mechanism is used to obtain different priorities of different items, and the long-term and short-term preferences of the user are combined to make more comprehensive recommendations, thereby improving the performance and accuracy of prediction.

[0039] The advantages of the additional aspects of the present application will be partially given in the following description, partially will become obvious from the following description, or will be understood through the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0040] The drawings accompanying the specification of the present application serve to provide a further understanding of the present application, and the illustrative embodiments of the present application and their descriptions serve to explain the present application, and do not constitute an improper limitation on the present application.

[0041] Figure 1 The method flowchart of the first embodiment.

[0042] Figure 2 The system structure diagram of the second embodiment. DETAILED DESCRIPTION

[0043] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0044] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0045] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0046] The overall concept proposed in this invention is as follows:

[0047] Session-based graph convolutional networks take a session sequence S and dwell time T as input, and then output the probability of all possible clicked items. As a score, the top k items with the highest scores are recommended to the user. The dwell time is incorporated into the conversation graph and used as the input to the graph convolutional network. The resulting node vector features are passed through a feature selection layer and used as the input to the target attention network. Then, the user's long-term preferences are obtained using a self-attention mechanism. Finally, the long-term and short-term preferences are combined to predict the user's preferences.

[0048] Example 1

[0049] This embodiment discloses a graph convolutional network session recommendation method with dwell time;

[0050] like Figure 1 As shown, the graph convolutional network session recommendation method with dwell time includes:

[0051] S11: Obtain the project and session sequence, and scale the dwell time in the session sequence;

[0052] In a conversation-based recommendation algorithm, V = {v1, v2, ..., v...} m} represents the set of all items involved in all sessions, and the user interaction sequence {v s,1 ,v s,2 ,......v s,n}, sorted by timestamp, where the interactive items involved in the interactive sequence S are represented as v. s,i ∈V, dwell time T={t1,t2,......,t n} represents the dwell time distribution of the session sequence S, t n ∈T means from item v s,n To Project v s,n+1 The dwell time, the purpose of session-based recommendation methods is to determine the dwell time based on historical interaction sequences v s,n Predict the user's next interaction event vs,n+1 .

[0053] The dwell time is calculated according to the timestamp, the unit is millisecond (ms), because the dwell time distribution is too sparse to affect the acquisition of user interest, so the dwell time is scaled before being merged into the session graph, two methods of normalization and inverse normalization are used to scale the dwell time, when the dwell time and the relationship strength are positively correlated (i.e. the longer the dwell time between two items, the stronger the relationship between them), the normalization method is used; when the dwell time and the relationship strength are negatively correlated (i.e. the shorter the dwell time between two items, the stronger the relationship between them), the inverse normalization method is used.

[0054] The formula of normalization is:

[0055]

[0056] The formula of inverse normalization is:

[0057]

[0058] In the formula (1) and formula (2), t max and t min are the maximum and minimum dwell time in the list. The scaling is performed according to the session order, so t max and t min change in different sessions.

[0059] S12: based on the session sequence after scaling the dwell time, a directed session graph is constructed, and the weight of the edge in the session graph is set according to the dwell time and the number of occurrences of the edge;

[0060] Each session interaction sequence S can construct a directed session graph G S = (V S , E S ), V is the item node, E is the edge in the session graph, in the session graph, v s,i ∈V represents an item clicked by the user in the session s, and each edge (v s,i-1 , v s,i ) represents that the user clicks v s,i-1 after clicking v s,i in the session; since many items may appear multiple times in the sequence, a standardized weight is assigned to each edge, and the calculation method is the number of occurrences of the edge / the out-degree of the starting node of the edge; the weight of the number of occurrences of the edge is represented as (w s,i-1 , w s,i ), and the scaled dwell time is multiplied by the number of occurrences; for example, from v s,i-1 to v s,iThe conversion between the two occurs three times, the first time with a dwell time of t'1, the second time with a dwell time of t'2, and the weight of the edge (w s,i-1 s,i ) is 3*(t'1+t'2), and the product of the dwell time and the number of occurrences represents the strength of the relationship between the two nodes.

[0061] S13: Extract features in the session graph through the graph convolution network, and screen repeated features by using a feature selection method based on a penalty term.

[0062] Embed each item v∈V into a unified vector space, and the node vector e∈R d represents the item v embedding vector learned through the graph convolution network, where d represents the dimension of the embedded node, and each session s is represented by an embedding vector s, which is composed of the node vectors in the session graph. The graph convolution network is very suitable for session-based recommendation, and can automatically extract features of the session graph, taking into account the rich node connections. The following is the learning process of the session graph node vector, and the update process of the node embedding in the graph is as follows:

[0063]

[0064] In formula (3), U j→i represents the information converted from item j to item i, and Ni represents the neighbor node set of item i. W j represents the weight matrix of the edge, represents the embedding of item j.

[0065]

[0066]

[0067] Formula (4) calculates the neighbor node information of v s,i , and formula (5) converts the neighbor node information into the embedding of the neighbor node. σ(·) is an activation function, and W a is a parameter matrix.

[0068]

[0069]

[0070]

[0071] respectively represent the forget gate, the input gate and the output gate. represents the information left and discarded from the upper layer, represents the information that needs to be updated, represents the final input new information, and b f ​is a bias term of the forget gate, b z is a bias term of the input gate, b r is a bias term of the output gate.

[0072]

[0073] In formula (9), the tanh(·) function is used to re-input information, and the upper layer information is generated is the candidate information in the finally generated new information, b p is a bias term.

[0074]

[0075]

[0076] Formula (10) represents that the information retained through the forget gate and the new information selected through the input gate are combined together to form the final new information. Formula (11) represents that part of the information in the final new information is selected through the output gate to form e s,i new embedding.

[0077] The features output by the graph convolution network have overfitting problems, and the feature selection layer is added to select the features based on the penalty term.

[0078] The L1 penalty term-based logistic regression model is used to retain the features with the same final recommendation value, and the L2 penalty term-based logistic regression model is used to optimize the remaining features. Specifically, if the feature weight is 1 in L1, a similar set is formed, the features with small differences in the weight in L2 and the weight of 0 in L1 in the set are divided into L1 equal weight features.

[0079] S14: Capture the long-term and short-term preferences of the user by using the attention mechanism, and recommend the items according to the long-term and short-term preferences of the user.

[0080] The target attention network and the self-attention mechanism are constructed, the target attention network is used to activate the user interest related to the target item, the self-attention mechanism is used to capture the long-term preference of the user, and the complex conversion between remote items is considered, and the attention scores between all items v i and each target item v i ∈V in the session s are further calculated as the ratings of the user to the items, and the items are recommended according to the rating results.

[0081] Embodiment two

[0082] The embodiment discloses a graph convolution network session recommendation system with dwell time;

[0083] As shown in the figure, the graph convolution network session recommendation system with dwell time comprises: Figure 2

[0084] The dwell time scaling module is configured to obtain items and session sequences, and scale dwell time in the session sequences;

[0085] The session graph construction module is configured to construct a directed session graph based on the session sequences after scaling the dwell time, and set the weight of the edge in the session graph according to the dwell time and the occurrence frequency of the edge;

[0086] The feature extraction and selection module is configured to extract features in the session graph through a graph convolution network, and screen repeated features by using a feature selection method based on a penalty term;

[0087] The item recommendation module is configured to capture long-term and short-term preferences of a user by using an attention mechanism, and recommend items according to the long-term and short-term preferences of the user.

[0088] Embodiment three

[0089] The purpose of the embodiment is to provide a computer-readable storage medium.

[0090] A computer-readable storage medium, which stores a computer program, the program being executed by a processor to implement the steps in the graph convolution network session recommendation method with dwell time as described in Embodiment 1 of the present disclosure.

[0091] Embodiment four

[0092] The purpose of the embodiment is to provide an electronic device.

[0093] An electronic device comprises a memory, a processor, and a program stored in the memory and executable on the processor, and the processor implements the steps in the graph convolution network session recommendation method with dwell time as described in Embodiment 1 of the present disclosure when executing the program.

[0094] The steps and methods involved in the above embodiments two, three and four correspond to Embodiment one, and the specific embodiments can be referred to the relevant description part of Embodiment one. The term "computer-readable storage medium" should be understood as including a single medium or multiple media of one or more instruction sets; it should also be understood as including any medium capable of storing, encoding or carrying instruction sets for execution by a processor and causing the processor to perform any method in the present application.

[0095] ​Those skilled in the art should understand that the modules or steps of the present application described above can be realized by general computer devices, or alternatively, they can be realized by program codes executable by the computer devices, so that they can be stored in the storage devices and executed by the computer devices, or they can be respectively manufactured into individual integrated circuit modules, or a plurality of modules or steps among them can be manufactured into a single integrated circuit module. The present application is not limited to any specific combination of hardware and software.

[0096] The specific embodiments of the present application described above with reference to the accompanying drawings are not intended to limit the protection scope of the present application, and those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A session recommendation method with graph convolution network and dwell time, characterized in that, an item and a session sequence are obtained, and the dwell time in the session sequence is scaled; a directed session graph is constructed based on the session sequence scaled by the dwell time, and the weight of the edge in the session graph is set according to the dwell time and the number of occurrences of the edge; features in the session graph are extracted through a graph convolution network, and repeated features are screened by a feature selection method based on a penalty term; user long-term and short-term preferences are captured by using an attention mechanism, and item recommendation is performed according to the user long-term and short-term preferences; Wherein, the directed conversation graph G is constructed S = (V S ,E S ), v s,i ∈V represents a project represents a node, and each edge (v s,i-1 , v s,i ) represents that the user clicks v s,i-1 after clicking v s,i in the conversation, and the normalized weight is set for each edge; the weight is calculated in the form of dwell time and occurrence times, indicating the strength of the relationship between two nodes; wherein the feature selection is based on a logistic regression model with an L1 penalty term, and the features with the same recommendation value are retained, and a logistic regression model with an L2 penalty term is used for optimization; the specific logic of the feature selection is that if the feature weight is 1 in L1, a similar set, their weight difference is not large in L2, and the weight is 0 in L1, the features in this set are divided into equal weights in L1.

2. The method of claim 1, wherein the graph convolutional network with dwell time session recommendation method is characterized by, The session sequence S = {v s,1 ,v s,2 ,......v s,n} is a collection of items ordered by timestamps, and the dwell time T = {t1, t2,......,t n} represents the dwell time distribution of the session sequence S, and t n ∈ T represents the dwell time from item v s,n to item v s,n+1 . 3.The method of claim 1, wherein, Normalization and denormalization methods are used to scale the dwell time. 4.The method of claim 1, wherein, The extraction method of the node feature vector in the graph convolution network includes the following processes: According to the session graph, the item feature vector is extracted; Based on the item feature vector, the neighbor information of the node is aggregated through the activation function; The retained feature information is obtained by using the forget gate, the input gate and the output gate; The candidate information in the retained feature information is obtained by using the activation function; The final new information is obtained by merging the information retained by the forget gate and the candidate information selected by the input gate; The output gate is used to select and output the new information obtained.

5. A graph convolutional network session recommendation system with dwell time, characterized by: It includes: The dwell time scaling module is configured to obtain an item and a session sequence, and scale the dwell time in the session sequence; The session graph construction module is configured to construct a directed session graph based on the session sequence scaled by the dwell time, and set the weight of the edge in the session graph according to the dwell time and the number of occurrences of the edge; The feature extraction and selection module is configured to extract features in the session graph through a graph convolution network, and screen repeated features by using a feature selection method based on a penalty term; The item recommendation module is configured to capture user long-term and short-term preferences by using an attention mechanism, and perform item recommendation according to the user long-term and short-term preferences; Wherein, the directed conversation graph G is constructed S = (V S , E S ), v s,i ∈V represents a project represents a node, and each edge (v s,i-1 , v s,i ) represents that the user clicks v s,i-1 after clicking v s,i in the conversation, and the normalized weight is set for each edge; the weight is calculated in the form of dwell time and occurrence times, indicating the strength of the relationship between two nodes; Wherein the feature selection is based on a logistic regression model with an L1 penalty term, and the features with the same recommendation value are retained, and a logistic regression model with an L2 penalty term is used for optimization; the specific logic of the feature selection is that if the feature weight is 1 in L1, a similar set, their weight difference is not large in L2, and the weight is 0 in L1, the features in this set are divided into equal weights in L1.

6. A computer-readable storage medium having stored thereon a program, characterized in that, The program is executed by the processor to realize the steps in the session recommendation method with graph convolution network and dwell time according to any one of claims 1-4.

7. An electronic device comprising a memory, a processor, and a program stored on the memory and executable on the processor, characterized in that, The processor executes the program to realize the steps in the session recommendation method with graph convolution network and dwell time according to any one of claims 1-4.

Images