Sequence recommendation method and system based on social network and implicit relationship and application

Abstract

The invention provides a sequence recommendation method and system based on a social network and an implicit relation and application, and relates to the technical field of personalized recommendation, and the method comprises the steps that an implicit user relation construction module constructs an implicit user relation according to the interest similarity between users; the implicit article relationship construction module is used for carrying out frequent item set mining on article interaction sequences of all users, calculating support degrees and confidence degrees among articles and constructing an implicit article relationship; a total heterogeneous graph generation module synthesizes the implicit user relationship, the implicit article relationship, the user social network and the interaction sequence of the user and the article to generate a total heterogeneous graph; a user-to-article interestingness prediction module predicts the interestingness of the user to the article according to the constructed total heterogeneous graph and the interaction sequence of the user and the article; and the article recommendation module performs sequential recommendation. According to the method, the interest similarity among the users is analyzed, and the implicit user relationship is established, so that the problems of sparse social relationships and non-uniform distribution in the social network are solved, and the recommendation effect of the system is improved.

Description

Technical Field

[0001] This application relates to the technical field of personalized recommendation, and more specifically, to a sequential recommendation method, system and application based on social network and implicit relationship. Background Art

[0002] With the rapid development of the Internet and the continuous growth of user behavior data, personalized recommendation has become one of the core technologies for information filtering and content distribution. The main goal of a recommendation system is to predict the content that a user may be interested in in the future based on the user's historical behavior and preferences, so as to improve user satisfaction and platform operation efficiency. In the prior art, typical personalized recommendation technologies include: (1) The Chinese invention patent with the publication number CN103886487A discloses a personalized recommendation method and system for a B2B platform on a Hadoop platform based on distributed technology. First, the invention is based on the distributed storage technology of Hadoop to centrally place, store and query various data such as website log files, product information, user information, etc. Secondly, the Hive service under the Hadoop platform is used to preprocess the data, and the recommendation algorithm is implemented using Map / Reduce. Then, Map / Reduce is used to implement information retrieval and text mining of text information, match the product information required by the user for inquiry and procurement, and obtain personalized recommendation information. Finally, the HBase service under the Hadoop platform is used to provide big data storage and query to improve the response of the website to recommended users.

[0003] (2) The Chinese invention patent with the publication number CN106384259A discloses a solution for a recommendation system that integrates social information, including: selecting the ratings of users for products within a period of time, using these ratings of users for these products, user friendship relationship information and the user's own characteristics as a data set: counting the total number of users and the total number of products, constructing a user-product true rating matrix, randomly selecting a part of the rating data as a training set and simultaneously constructing a user friendship matrix and a user own characteristic matrix; calculating the similarity between a user and his friends; adopting a latent semantic model to generate a user feature matrix and a product feature matrix, and multiplying the matrices to obtain a predicted rating; calculating the root mean square error RMSE between the predicted rating and the true rating, and adding the similarity between users as the weight of the social information for the error. The gradient descent algorithm is used to iteratively update the user features and product features.

[0004] However, the social graphs in the real world often show a long-tail distribution, that is, most users have very few social connections, resulting in insufficient transmission of recommendation information based on the social network, thus affecting the model's ability to model social relationships. Moreover, the existing sequential recommendation models mainly focus on the positional relationship of items in the sequence and do not fully consider the purchase logic between items. Summary of the Invention

[0005] To solve the above problems, the technical solution adopted in this application is a sequential recommendation system based on social networks and implicit relationships, including: Implicit user relationship construction module: used to construct implicit user relationships according to the interest similarity between users; Implicit item relationship construction module: used to calculate the support and confidence between items by performing frequent item set mining on the item interaction sequences of all users, and then construct implicit item relationships; Total heterogeneous graph generation module: used to comprehensively generate a total heterogeneous graph containing all the above interaction relationships based on implicit user relationships, implicit item relationships, user social networks, and user-item interaction sequences; User interest prediction module for items: used to predict the user's interest in items according to the constructed total heterogeneous graph and the user-item interaction sequence; Item recommendation module: used to perform sequential recommendations based on the user's interest in items.

[0006] Optionally, the implicit user relationship construction module includes: Item-based user interest similarity calculation unit, which calculates the interest similarity of user and in item i according to the following formula: ; In the formula, are the ratings of user and in item i respectively, and is the interest similarity of user and in item i; User cumulative interest similarity calculation unit, which calculates the cumulative interest similarity of user according to the following formula: ; In the formula, represents the set of items rated by both user and , and represents the cumulative interest similarity of user ; Implicit user relationship construction unit, which defines , and only retains the top M users sorted in descending order according to the value. The implicit user relationship of user is calculated according to the following formula: ; In the formula, Represents a user where N represents the interest similarity threshold and M represents the number of implicit neighbors of each user.

[0007] Optionally, the calculation of the interest similarity threshold includes defining a set of interest similarities between users as ={ , ,…, },user The number of purchases is , the number of purchases of all users is , interest similarity threshold The calculation is based on the following formula: ; Where, is the collection of interest similarity The mean of is the collection of interest similarity The standard deviation of max{A u} represents the maximum number of shopping times; The calculation of the number of implicit neighbors for each user consists of Represents a user The number of categories of goods purchased, the number of categories of items purchased by each user is , the number of neighbors M is calculated according to the following formula: ; In the formula, max{A u} indicates the maximum number of product categories.

[0008] Optionally, the implicit item relationship building module includes: The item set support calculation unit is calculated according to the following formula and Itemset of two items Support : ; Where, Represents a transaction database Contains The number of transactions; The item set confidence calculation unit generates association rules according to the following formula : ; Where, Represents association rules confidence level; Implicit item relationship building unit, limited , , and only keep the first items sorted in descending order of value. The item hidden relationship is calculated based on the following formula: ; In the formula, S represents the item set support threshold, represents the implicit item relationship of item .

[0009] Optionally, the total heterogeneous graph generation module includes: Taking and to represent the sets of users and items respectively, define the user-item bipartite graph as , where the edge , and sorted by time, the items interacted by user can form the item interaction sequence of , where each element represents the item interacted at time . For the user-user relationship, define as the social network graph, where the edge ; Combine the user-item bipartite graph and the social network graph to obtain the heterogeneous graph ; Extract the user hidden relationship and the item hidden relationship, and define the user hidden relationship graph and the item hidden relationship graph as , the social network graph is extended to , define the item association graph as , and the total heterogeneous graph is extended to .

[0010] Optionally, the user interest prediction module for items includes an item interaction sequence-based user interest prediction module and a social network-based user interest prediction module.

[0011] Optionally, the item interaction sequence-based user interest prediction module includes: An item aggregation unit based on the implicit item relationship, which calculates the item aggregation of item according to the following formula: ; In the formula, the item interaction sequence of user is , for item , is the neighbor of item based on the item hidden relationship, yes The neighbor set of is the weight matrix, is bias; user The item interaction sequence after item aggregation is: ; Item aggregation unit based on item interaction sequence: For item aggregation ,by Indicates the position in the item interaction sequence Position embedding, and is the trainable weight matrix, is the embedding dimension, is the scaling factor, and the importance weight of historical items in the item interaction sequence is calculated by the following formula: ; Where, Indicates the position in the item interaction sequence Position embedding, Represents historical item aggregation Position in item aggregation Before; is a trainable weight matrix, item The items are aggregated into: ; user The item interaction sequence after item aggregation based on the item interaction sequence is: ; User aggregation unit based on social network and implicit user relationship: set up is a user The set of neighbors in a social network, For the neighbor set in the implicit user relationship, the total neighbor set of the user is defined as ,make express activation function, 、 and is the trainable weight matrix, 、 and is the bias vector, is the concatenation operation between vectors, Represents a user The feature vector of the neighbor user, Indicates each At the weight during user aggregation is calculated based on the following formula: ; user aggregation is calculated based on the following formula: ; Aggregation unit of users based on item time limit: Given user 's item interaction sequence , set the time limit , for the item interacted by the user at time, find other users who also interact with within the time window as for item under the time window limit, is defined as: ; For the item interaction sequence under the time window limit, is defined as: ; is the set of neighbors of user based on item time limit, is 's total neighbor set, user aggregation is calculated based on the following formula: ; In the formula, GAT() represents the graph neural attention network. Based on this, , for the item sequence , the neighbors of the user based on the item time limit are: ; User interest calculation unit from the perspective of item interaction sequence: Based on the aforementioned unit operations, predict the score of user for the next item from the perspective of the item interaction sequence of user , and are trainable weight matrices, is the bias vector: ;

[0012] ; ; wherein, GRU represents a gated recurrent unit, and ReLU is an activation function.

[0013] Optionally, the user interest prediction module from the social network perspective includes: An aggregation unit based on the item interaction sequences of the user's neighbors: After aggregating items based on implicit item relationships, the item interaction sequence of the user is , represents a trainable weight matrix, and the total neighbor set of the user is , and the aggregation of the item interaction sequences based on the user's neighbors is: ; ; wherein, represents the item interaction sequences based on the user's neighbors; A user aggregation unit based on user relationships: For the user , is the total neighbor set of is a trainable weight matrix, is the bias vector, and the neighbor aggregation of the user is as follows: ; A user interest calculation unit from the social network perspective: Predict the score of the user for the next item from the user's social network perspective according to the following formula, where and are trainable weight matrices, and are bias vectors: ; .

[0014] The present application also provides a sequential recommendation method based on social networks and implicit relationships, which is carried out by using the sequential recommendation system based on social networks and implicit relationships as described in any of the foregoing, including: Construction of implicit user relationships: constructing implicit user relationships according to the interest similarity between users; Construction of implicit item relationships: by performing frequent item set mining on the item interaction sequences of all users, calculating the support and confidence between items, and then constructing implicit item relationships; Generation of the total heterogeneous graph: integrating implicit user relationships, implicit item relationships, user social networks, and user-item interaction sequences to generate a total heterogeneous graph containing all the above interaction relationships; Prediction of user interest in items: predicting the user's interest in items according to the constructed total heterogeneous graph and the user-item interaction sequence; Item recommendation: performing sequential recommendation based on the user's interest in items.

[0015] The present application also provides a sequential recommendation application based on social networks and implicit relationships, which is carried out by using the sequential recommendation system based on social networks and implicit relationships as described in any of the foregoing, including applications to e-commerce platforms, online video and audio platforms, social media, content information applications, and online education platforms.

[0016] The beneficial effects of the sequential recommendation method, system, and application based on social networks and implicit relationships provided by the present application are as follows: (1) By analyzing the interest similarity between users, implicit user relationships are established to alleviate the problems of sparse and unevenly distributed social relationships in social networks, thereby improving the recommendation effect of the system; (2) Using frequent item set mining technology to identify potential purchase logic relationships between items, thereby establishing implicit item relationships, and further improving the accuracy of model recommendations; a sequential recommendation model that fully considers the order of purchase logic between items can improve the robustness and performance of the model in scenarios where social networks are sparse and the mining of item purchase logic is insufficient; (3) Considering both the long-term preferences and short-term interest changes of users, accurately capturing the dynamic behavior patterns of users, and further improving the timeliness and accuracy of recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art.

[0018] Figure 1 is the structural block diagram of the sequential recommendation system based on social networks and implicit relationships provided by the embodiments of the present application; Figure 2It is a block diagram of the implicit user relationship construction module provided by an embodiment of the present application; Figure 3 It is a block diagram of the implicit item relationship construction module provided by an embodiment of the present application; Figure 4 It is a block diagram of the overall heterogeneous graph generation module, the user's interest degree prediction module for items, and the item recommendation module provided by an embodiment of the present application; Figure 5 It is the overall heterogeneous graph provided by an embodiment of the present application. Detailed implementation manners

[0019] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0020] Embodiment 1 As Figure 1 shown, the present application provides a sequential recommendation system based on social networks and implicit relationships, including: Implicit user relationship construction module: used to construct implicit user relationships according to the interest similarity between users; by analyzing the interests of two different users in the same item, the interest similarity between the two users in this item is obtained. Similarly, by statistically calculating the interest similarity between two users in all items they have interacted with together, the interest similarity between the two users can be obtained. An implicit relationship between users is constructed according to the interest similarity between users.

[0021] Implicit item relationship construction module: used to perform frequent item set mining on the item interaction sequences of all users, calculate the support and confidence between items, and then construct implicit item relationships; Overall heterogeneous graph generation module: used to comprehensively generate an overall heterogeneous graph including all the above interaction relationships based on implicit user relationships, implicit item relationships, user social networks, and the interaction sequences between users and items; User's interest degree prediction module for items: used to predict the user's interest degree in items according to the constructed overall heterogeneous graph and the interaction sequences between users and items; Item recommendation module: used to perform sequential recommendation based on the user's interest degree in items.

[0022] Aiming at the problems of sparse social relationships and uneven distribution in traditional recommendation systems, by deeply analyzing the interest similarity between users, that is, calculating the similarity of the user's historical behavior sequences, implicit user relationships are constructed between users who have no direct social connection, as Figure 2 shown.

[0023] The following is a specific description of each unit: Item-based user interest similarity calculation unit: Specifically, if the ratings of two users on the same item are similar, it can be inferred that their interests are also relatively close. Assume that user and 's ratings on item i are and respectively, then the interest similarity and of user on item i is calculated as follows: <x ; User cumulative interest similarity calculation unit: By accumulating the interest similarities of user on all co-rated items, the cumulative interest similarity and of user can be obtained: ; where represents the set of items that have been rated by both user and .

[0024] Implicit user relationship construction unit: To obtain a closer hidden user relationship, we require , and only keep the top users sorted in descending order of value. Then the implicit user relationship of user is: ; And the calculation methods of parameters N and M are as follows: In an e-commerce platform, the user interest similarity is significantly affected by the activity level of their shopping behavior. Therefore, when calculating the interest similarity threshold N, we introduce the user activity (purchase frequency) as a weighting factor. Specifically, define the set of interest similarities between users as ={ , ,…, }, the number of purchases of user is , and the set of the number of purchases of all users is . Then the calculation method of the interest similarity threshold is: ; where is the set of interest similarities The mean value of is the set of interest similarities The standard deviation of

[0025] For the number M of implicit neighbors of each user, two factors are comprehensively considered: the more categories of products a user has purchased historically, the more extensive their interests are, and more neighbors should be retained. Let represent the user The number of product categories purchased, and the set of the number of item types purchased by each user is Then the calculation method of the neighbor number M is: ; In this way, the present invention can significantly alleviate the problem of sparse social relationships, expand the source of users' social information, and thus improve the integrity of user embedding representation and the performance of the recommendation system.

[0026] Such as Figure 3 shown, the implicit item relationship construction module includes: an item set support calculation unit, an item set confidence calculation unit, and an implicit item relationship construction unit. In actual situations, if two items frequently co-occur in a user's shopping history, it indicates that there may be a certain purchase association between them. For example, a classic case is This frequent item set, which indicates that these two products tend to be purchased simultaneously, has important reference value for predicting user purchase behavior. This module uses frequent item set mining technology to analyze the user's historical behavior sequence, and mines and constructs the implicit association relationship between items.

[0027] The following is a specific description of each unit: Item set support calculation unit: The item set support measures the frequency of occurrence of a certain item set in the transaction database. Let and be two items, then the support of the item set is . represents the number of transactions in the transaction database that contain ; Item set confidence calculation unit: <> Confidence measures the probability that appears in the case where also appears, and is used to generate association rules . The confidence is defined as follows: ; In the formula, represents the association rule Confidence; Implicit item relationship construction unit: To obtain more effective hidden item relationships, it is defined that , , and sorted in descending order according to the value, and the top K items are taken as the items having hidden relationships with item : ; In the formula, S represents the item set support threshold, represents the implicit item relationship of item , and S and K are positive integers greater than 0.

[0028] The value of S is , where represents the average support of all binomial item sets, represents the standard deviation of the support of all binomial item sets. The value of K is , where L represents the size of the set .

[0029] As shown in Figure 1 and Figure 4 , the specific description of the total heterogeneous graph generation module is as follows: Let and represent the sets of users and items respectively. We define the user-item bipartite graph as , where the edge , and sorted by time, the items interacted by user can form the item interaction sequence of , where each element represents the item interacted by at time . For the user-user relationship, we define as the social network graph, where the edge . Combining the user-item bipartite graph and the social network graph, we obtain the following heterogeneous graph .

[0030] In addition, as in the operations of the implicit user relationship construction module and the implicit item relationship construction module, we further extract the user hidden relationship and the item hidden relationship, and define the user hidden relationship graph and the item hidden relationship graph as . Then the social network graph can be extended to , define the item association graph as , then the total heterogeneous graph can be extended to , as shown in Figure 5 .

[0031] As Figure 4 shown, the user interest prediction module for items includes a user interest prediction module from the perspective of item interaction sequences and a user interest prediction module from the perspective of social networks.

[0032] Among them, the user interest prediction module from the perspective of item interaction sequences includes an item aggregation unit based on implicit item relationships, an item aggregation unit based on item interaction sequences, a user aggregation unit based on social networks and implicit user relationships, a user aggregation unit for neighbors based on item time limits for users, and a user interest calculation unit (from the perspective of item interaction sequences).

[0033] The following is a specific description of each unit: Item aggregation unit based on implicit item relationships: Suppose the item interaction sequence of user is , for item , is the neighbor of item based on the hidden relationship of the item, is 's neighbor set, is the weight matrix, is the bias. Then the item aggregation of item is as follows: ; Then the item interaction sequence of user after item aggregation is: ; Item aggregation unit based on item interaction sequences: The user's choice of the next item is not only affected by the most recently interacted item but also by its complete historical behavior sequence. Therefore, when performing item aggregation, it is necessary to comprehensively consider the user's short-term interests and long-term preferences to more accurately model their behavior patterns. Inspired by the excellent performance of the self-attention mechanism in prediction tasks, this study uses a masked self-attention mechanism combined with item position information to aggregate the information of the item interaction sequence, thereby enhancing the item representation. The item interaction sequence used here is the item interaction sequence after item aggregation in the item aggregation unit based on implicit item relationships. .

[0034] For item , let be the position embedding at position in the interaction sequence, and are trainable weight matrices, is the embedding dimension, is the scaling factor. Then, the importance weight of historical items in the item interaction sequence is calculated by the following formula: ; In the formula, represents the position embedding of the position in the item interaction sequence, represents that the position of the historical item aggregation is before the item aggregation ; Let be a trainable weight matrix. Then, the item aggregation of item is: ; Then, the item interaction sequence of user after aggregation based on the item interaction sequence is: ; User aggregation unit based on social network and implicit user relationship: This unit uses a graph attention neural network (GAT) to further aggregate information on social relationships and implicit user relationships to optimize user representation. Specifically, let be the set of neighbors of user in the social network, and be its set of neighbors in the implicit user relationship. Then, the total neighbor set of the user is defined as . Let be the activation function, , and be trainable weight matrices, , and be bias vectors, be the concatenation operation between vectors, represents the weight of each during user aggregation . Then, is expressed as: ; The user aggregation is expressed as: ; User aggregation unit for neighbors based on item time limit: Given the item interaction sequence of user , a time limit is set. For the item interacted by the user at time , the time window Find other users who also interact with as For an item in a time window The user relationship under the limit is defined as: ; Then For the item interaction sequence in a time window The user relationship under the limit is defined as: ; is the set of neighbors of user based on the item time limit, . Similar to the user aggregation strategy of the user aggregation unit based on the social network and implicit user relationships, assume is the total set of neighbors, then the user aggregation of is expressed as: ; In the formula, GAT() represents the graph neural attention network. Based on this, . Finally, for the item sequence , the neighbors of the user based on the item time limit are: .

[0035] User interest calculation unit (from the perspective of item interaction sequence): Based on the aforementioned unit operations, predict the score of user for the next item from the perspective of the item interaction sequence of user , and are trainable weight matrices, and are bias vectors: ; ; ; In the formula, GRU represents the gated recurrent unit, and ReLU is the activation function.

[0036] The user interest prediction module from the social network perspective includes an aggregation unit for the item interaction sequences of user neighbors, a user aggregation unit for user relationships, and a user interest calculation unit (from the social network perspective).

[0037] The following is a specific description of each unit: Aggregation unit of item interaction sequences based on user neighbors: According to the description of the item aggregation unit based on implicit item relationships, after item aggregation based on implicit item relationships, the item interaction sequence of the user is . From the user aggregation unit based on the social network and implicit user relationships, it can be known that the total neighbor set of the user is . Then, the aggregation of item interaction sequences based on user neighbors is defined as: ; ; User aggregation unit of user relationships: For user , is the total neighbor set of, is a trainable weight matrix, is a bias vector. Then, the neighbor aggregation of user is as follows: ; User interest calculation unit (from the perspective of the social network): [[ID=*46]]Based on the operations of the foregoing units, predict the score of user for the next item from the perspective of the user's social network in the following manner , where and are trainable weight matrices, and are bias vectors: ; ; Item recommendation module, comprehensively calculates the total interest degree of the user for the item by combining the interest degree predicted from the perspective of the interaction sequence between the user and the item and the interest degree predicted from the perspective of the social network, and generates a sorted set arranged in descending order of interest degree.

[0038] The specific description is as follows: Integrate the content of the module, and predict the scores of items from the perspectives of the user's item interaction sequence and the social network respectively. Finally, predict the total interest degree score of user for item as: ; Note: There seems to be an error in the original text at line 46 where the text is cut off. I've translated it as best as possible based on the context.Finally, sort the items that the user will interact with from top to bottom according to the interest score.

[0039] Embodiment 2 The present application also provides a sequential recommendation method based on social networks and implicit relationships, which is carried out by using the sequential recommendation system based on social networks and implicit relationships as described in any of the foregoing, including: Constructing implicit user relationships: constructing implicit user relationships according to the interest similarity between users; Constructing implicit item relationships: by performing frequent itemset mining on the item interaction sequences of all users, calculating the support and confidence between items, and then constructing implicit item relationships; Generating a total heterogeneous graph: comprehensively combining implicit user relationships, implicit item relationships, user social networks, and user-item interaction sequences to generate a total heterogeneous graph containing all the above interaction relationships; Predicting the user's interest in items: predicting the user's interest in items according to the constructed total heterogeneous graph and the user-item interaction sequences; Item recommendation: performing sequential recommendation based on the user's interest in items.

[0040] Embodiment 3 The present application also provides a sequential recommendation application based on social networks and implicit relationships, which is carried out by using the sequential recommendation system based on social networks and implicit relationships as described in any of the foregoing, including applications to e-commerce platforms, online video and audio platforms, social media, content information applications, and online education platforms.

[0041] For an e-commerce platform, the "items" in Embodiment 1 represent the goods provided by the e-commerce platform. For an online video and audio platform, the "items" in Embodiment 1 represent the videos or audios provided by the online video and audio platform.

[0042] For social media, the "items" in Embodiment 1 represent the content posted by users on social media.

[0043] For content information applications, the "items" in Embodiment 1 represent the information provided by the content information applications.

[0044] For an online education platform, the "items" in Embodiment 1 represent the courses provided by the online education platform.

[0045] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included within the protection scope of the present application.

Claims

1. A sequential recommendation system based on social networks and implicit relationships, characterized in that Including: Implicit user relationship construction module: used to construct implicit user relationships based on the interest similarity between users; Implicit item relationship construction module: used to perform frequent itemset mining on the item interaction sequences of all users, calculate the support and confidence between items, and then construct implicit item relationships; Total heterogeneous graph generation module: used to synthesize implicit user relationships, implicit item relationships, user social networks, and user-item interaction sequences to generate a total heterogeneous graph containing all the above interaction relationships; User interest prediction module for items: used to predict the user's interest in items based on the constructed total heterogeneous graph and the user-item interaction sequence; Item recommendation module: used to perform sequential recommendations based on the user's interest in items.

2. The sequential recommendation system based on social network and implicit relationship according to claim 1, characterized in that: The implicit user relationship construction module includes: An item-based user interest similarity calculation unit calculates the interest similarity of user and in item i according to the following formula: ； Wherein, are respectively the scores of user and on item i, is the interest similarity between user and on item i; The user cumulative interest similarity calculation unit calculates the cumulative interest similarity of the user according to the following formula: The cumulative interest similarity is as follows: ； In the formula, represents the set of items that have been rated by both the user and ; represents the cumulative interest similarity of the user ; ​ Implicit user relationship construction unit, defined as , and only retain the top M users sorted in descending order according to the value, and the implicit user relationship of the user is calculated based on the following formula: ； In the formula, represents the implicit user relationship of the user , N represents the interest similarity threshold, and M represents the number of implicit neighbors of each user.

3. The sequential recommendation system based on social network and implicit relationship according to claim 2, characterized in that: The calculation of the interest similarity threshold includes defining the set of interest similarities between users as ={ , ,…, }, the number of shopping times of user is , and the set of the number of shopping times of all users is . The interest similarity threshold is calculated according to the following formula: ； Wherein, is the set of interest similarities mean value of, is the set of interest similarities standard deviation of, max{A u} represents the maximum number of shopping times; The calculation of the number of implicit neighbors for each user includes making represent the number of product categories purchased by user . The set of the number of item types purchased by each user is , and the calculation of the number of neighbors M is carried out according to the following formula: ； where max{A u} represents the maximum number of commodity categories.

4. The sequential recommendation system based on social networks and implicit relationships according to claim 1, characterized in that: The implicit item relationship construction module includes: The item set support calculation unit calculates according to the following formula and the item set of two items support : ； In the formula, represents the number of transactions contained in the transaction database; The item set confidence calculation unit generates association rules according to the following formula :[[]]END]] ； In the formula, represents the confidence of the association rule ; Implicit item relationship construction unit, defined as , , and only keep the top items sorted in descending order of value. The item hiding relationship of item is calculated based on the following formula: ； Where S represents the item set support threshold, represents an item and its implicit item relationship.

5. The sequential recommendation system based on social networks and implicit relationships according to claim 1, characterized in that: The total heterogeneous graph generation module includes: Let and represent the sets of users and items respectively. Define the user-item bipartite graph as where the edge and sorted by time. The items interacted by user can form an item interaction sequence where each element represents the item interacted at time . For the user-user relationship, define as the social network graph where the edge . Combine the user-item bipartite graph and the social network graph to obtain the heterogeneous graph ; ; Extract the hidden relationships of users and items, and define the user hidden relationship graph and the item hidden relationship graph as , the social network graph Expand it to , define the item association graph as , expand the total heterogeneous graph to .

6. The sequential recommendation system based on social network and implicit relationship according to claim 1, characterized in that: The user interest prediction module for items includes an item interaction sequence perspective user interest prediction module and a social network perspective user interest prediction module.

7. The sequential recommendation system based on social networks and implicit relationships according to claim 6, wherein: The item interaction sequence perspective user interest prediction module includes: An item aggregation unit based on implicit item relationships calculates the item aggregation of an item according to the following formula of the item as follows: ； where the user 's item interaction sequence is , for item , is the neighbor of item based on the item hiding relationship, is 's neighbor set, is the weight matrix, is the bias; User The item interaction sequence of ; Item aggregation unit based on item interaction sequences: For item aggregation , use to represent the positional embedding at position in the item interaction sequence, and are trainable weight matrices, is the embedding dimension, is the scaling factor. The importance weights of historical items in the item interaction sequence are calculated by the following formula: ； In the formula, represents the position in the item interaction sequence, represents the historical item aggregation where the position of is before the item aggregation; is a trainable weight matrix, item of items is aggregated as: ； User The item interaction sequence after item aggregation based on the item interaction sequence is as follows: ； User aggregation unit based on social network and implicit user relationships: Let be the set of neighbors of the user in the social network, be the set of neighbors of the user in the implicit user relationship. Define the total neighbor set of the user as . Let denote the activation function, , and be trainable weight matrices, , and be bias vectors, be the concatenation operation between vectors, denote the feature vector of the neighbor user of the user , denote the weight of each in the user aggregation. is calculated based on the following formula: ； User aggregation Calculated based on the following formula: ； Aggregation unit for neighbors of users based on item time limits: Given user Item interaction sequence , set a time limit , for users in Items that you interact with all the time , in the time window Found also with Other users who interact as For items In the time window The user relationship under the restriction is defined as: ； For the item interaction sequence In the time window The user relationship under the limit is defined as: ； is the user item-based time-limited neighbor set, is the total neighbor set of, user aggregation of calculated based on the following formula: ； wherein, GAT() represents a graph neural attention network, based on which, , for the item sequence , the user's neighbors based on the item time limit are: ; User interest calculation unit from the perspective of item interaction sequences: Based on the foregoing unit operations, predict the user's rating of the next item from the perspective of the user's item interaction sequence as follows : where , and are trainable weight matrices, is a bias vector: ； ； ； Wherein, GRU represents a gated recurrent unit, and ReLU is an activation function.

8. The sequential recommendation system based on social networks and implicit relationships according to claim 6, characterized in that: The social network perspective user interest prediction module includes: Aggregation unit for item interaction sequences based on user neighbors: After item aggregation based on implicit item relationships, the user 's item interaction sequence is , represents a trainable weight matrix, and the user 's total neighbor set is , and the item interaction sequence aggregation based on the user's neighbors is as follows: ； ； In the formula, represents the item interaction sequence based on the user's neighbors; User aggregation unit based on user relationships: For a user , is the total neighbor set of, is a trainable weight matrix, is a bias vector, the user neighbor aggregation is as follows: ； Social network perspective user interest calculation unit: Predict the user's rating of the next item from the perspective of the user's social network according to the following formula of the rating , where and are trainable weight matrices, and are bias vectors: ； 。 9. A sequential recommendation method based on social networks and implicit relationships, characterized in that: Using the sequential recommendation system based on social network and implicit relationship according to any one of claims 1-8, including: Implicit user relationship construction: constructing implicit user relationships based on the interest similarity between users; Implicit item relationship construction: performing frequent itemset mining on the item interaction sequences of all users, calculating the support and confidence between items, and then constructing implicit item relationships; Total heterogeneous graph generation: synthesizing implicit user relationships, implicit item relationships, user social networks, and user-item interaction sequences to generate a total heterogeneous graph containing all the above interaction relationships; User interest prediction for items: predicting the user's interest in items based on the constructed total heterogeneous graph and the user-item interaction sequence; Item recommendation: performing sequential recommendations based on the user's interest in items.

10. Sequential recommendation application based on social network and implicit relationships, characterized in that: Using the sequential recommendation system based on social network and implicit relationship according to any one of claims 1-8, including applications to e-commerce platforms, online video and audio platforms, social media, content information applications, and online education platforms.

Images