Interest group recommendation method based on double-intention contrastive learning in event social network

Abstract

The application discloses a kind of event social network based on double-intention contrast learning's interest group recommendation method and device, it is related to event social network technical field, the method includes: constructing online interest group hypergraph, obtains the online user representation and online interest group representation of fusion social intention;Offline event hypergraph is constructed, and offline user representation and offline interest group representation of offline intention perception are obtained;Enhance online interest group representation and offline interest group representation, obtain online enhanced interest group representation and offline enhanced interest group representation;After aligning respectively to online user representation, offline user representation, online enhanced interest group representation and offline enhanced interest group representation, obtain double-intention user representation and double-intention interest group representation;One of the target user and interest group is obtained, and the matching degree determined according to double-intention user representation and double-intention interest group representation determines target interest group.The application solves the problem that interest group recommendation is not accurate in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of event social networks, in particular to an interest group recommendation method and device based on double-intention contrast learning in an event social network. BACKGROUND

[0002] Event-based social networks (EBSNs), such as Meetup and Douban, integrate online social connections with offline event participation, providing a new type of social experience for user interaction. In such platforms, interest groups are online communities established by users around common interests, and they not only carry out long-term interactions among users, but also serve as the core carriers for initiating and organizing offline events. Interest group recommendation, as a key mechanism for matching user preferences with group resources, is crucial for improving user activity and platform value.

[0003] Existing mainstream recommendation methods based on GNN mainly model online multi-entity interactions and cannot capture high-order dependency relationships between online and offline behaviors. For example, a user initially joins an "indoor fitness interest group" to obtain online training methods and dietary advice. Inspired by the sharing within the group, he signs up for an offline "city marathon experience event". In the offline event, he develops a stronger interest in running. After the event, he joins a "marathon training interest group", further expanding his online social and learning circle. As can be seen, participating in an online interest group drives offline event participation, and offline event participation promotes online interest conversion. Existing research cannot capture this high-order dependency relationship between online and offline behaviors, resulting in a one-sided representation of user multi-dimensional behavior patterns, which limits the understanding of the user's real behavior patterns by the recommendation system.

[0004] Secondly, most existing intention modeling research relies on online behavior to infer user interest or behavior motivation, but ignores the key role of the intention behind the user's offline behavior in interest group recommendation. For example, a user joins an "acoustic guitar interest group" online, and his intention may be to "improve his guitar playing skills", based on which the system will recommend interest groups related to the guitar, such as advanced guitar courses and instrument evaluation interest groups. However, when the user subsequently participates in an offline "independent music and dance fusion performance" event, his motivation may have shifted from simply learning guitar to cross-border cooperation, stage performance, and artistic expression. If the recommendation system still only pushes based on online behavior intention, it will ignore the dynamic changes in the user's current interest and potential expansion direction, resulting in a deviation between the recommended content and the user's interest.

[0005] Thirdly, the existing model extracts the features of the interest group mainly depending on the static attributes such as interest group ID code and category label, and does not effectively fuse the semantic information of the events in the group. For example, in the recommendation of the "parent-child education" interest group, the user has participated in the "3-6 year-old children's picture book reading guidance" offline event, and the potential demand may be "early reading ability cultivation of young children". If the category labels of certain interest groups are also "parent-child education", but the recent events in the group are mainly "adolescent parent-child communication skill lectures", there is a clear mismatch between the semantic direction and the user's demand; while another group with the same label, the events in the group are mainly "picture book extension handicraft activities" and "children's reading habit cultivation workshop", the semantic information is highly consistent with the user's demand. The existing model fails to mine the semantic differences of such events, resulting in the inability to achieve accurate matching between the interest group and the user's real demand.

[0006] Therefore, the current interest group recommendation is difficult to depict the high-order dependency relationship between online and offline behaviors, ignores the important role of offline behavior intention in capturing the change of user interest, and the representation of the interest group is static and lacks event semantic information, thereby causing the problem of inaccurate recommendation. SUMMARY

[0007] Therefore, the current interest group recommendation is difficult to depict the high-order dependency relationship between online and offline behaviors, ignores the important role of offline behavior intention in capturing the change of user interest, and the representation of the interest group is static and lacks event semantic information, thereby causing the problem of inaccurate recommendation.

[0008] In one aspect, the application provides an interest group recommendation method based on double-intention contrast learning in an event social network, which comprises the following steps:

[0009] According to the online interaction relationship between the user and the interest group, an online interest group hypergraph is constructed, and a hypergraph neural network is used for representation learning of the online interest group hypergraph to obtain online user representation and online interest group representation fused with social intention;

[0010] According to the event held by the interest group and the event participation data of the members, an offline event hypergraph is constructed, and a hypergraph neural network is used for representation learning of the offline event hypergraph to obtain offline user representation and offline interest group representation perceived by offline intention;

[0011] The online interest group representation and the offline interest group representation are enhanced by the interaction relationship between the interest group and the event, to obtain online enhanced interest group representation and offline enhanced interest group representation;

[0012] The online user representation, the offline user representation, the online enhanced interest group representation and the offline enhanced interest group representation are aligned respectively to obtain double-intention user representation and double-intention interest group representation fused with double intention;

[0013] The target interest group is determined according to the matching degree determined by the double-intention user representation and the double-intention interest group representation, so as to recommend the target interest group.

[0014] Further, the interest group recommendation method based on double-intention contrast learning in the event social network, wherein the step of constructing an online interest group hypergraph according to the online interaction relationship between the user and the interest group, and performing hypergraph neural network representation learning on the online interest group hypergraph to obtain the online user representation and the online interest group representation fused with the social intention comprises:

[0015] The expression of the online interest group hypergraph is:

[0016] ;

[0017] The expression of the online user representation and the online interest group representation fused with the social intention is:

[0018] ;

[0019] Wherein, , U and G are a user set, an interest group set, is a hyperedge set, is an adjacency matrix of the online interest group hypergraph, M , K respectively represent the sizes of the user set and the interest group set, and when the node , is the member set of the interest group , , , respectively represent the user initial feature embedding matrix and the interest group initial feature embedding matrix, is a hypergraph neural network parameter, represents a classic hypergraph neural network, and the calculation formula is wherein, represents the feature embedding matrix of the i-th layer, represents the weight matrix of the i-th layer, represents the adjacency matrix, and respectively represent the degree matrix of the node and the hyperedge.

[0020] ​​Furthermore, the interest group recommendation method based on dual-intent contrastive learning in the aforementioned event social network includes the following steps: constructing an offline event hypergraph based on event data organized by interest groups and event participation data, and obtaining offline user representations and offline interest group representations for offline intent perception through hypergraph neural network representation learning of the offline event hypergraph.

[0021] The expression for the offline event hypergraph is:

[0022] ;

[0023] The expressions for offline user representation and offline interest group representation are:

[0024] ;

[0025] ;

[0026] in, Represents a set of users and event nodes. Let each superedge be a set of superedges. Corresponding to an interest group, Interest group member nodes and the event nodes they participate in are connected via hyperedges. This represents the adjacency matrix of the offline event hypergraph, where M, N, and K represent the sizes of the user set, event set, and interest group set, respectively. When the node... hour, ,in, For interest groups Member set, for A set of events in which members participate. The number of layers in which the message modules are stacked. l This is the layer index for the message module. , Represents user node The set of associated hyperedges Represents user node Associated hyperedge Embedded, ,in This indicates element-wise multiplication. This represents a trainable weight matrix. This indicates a splicing operation. , , Indicates interest group The set of members, Indicates interest group members Embedded, Indicates an event Embedded, denotes a node aggregation function, which is an attention aggregation mechanism, d denotes the embedding dimension.

[0027] Further, the interest group recommendation method based on double-intention contrast learning in the event social network, wherein the step of respectively enhancing the online interest group representation and the offline interest group representation through the interest group and event interaction relationship includes:

[0028] The interest group and event interaction relationship is represented as a bipartite graph

[0029] The interest group embedding is spliced with the event embedding , and the adjacency matrix is input into a graph neural network to obtain the interest group representation:

[0030]

[0031] The initial embedding is passed through L convolutional layers, and the embedding obtained at each layer is averaged to obtain the final interest group representation enhanced by events

[0032]

[0033] The interest group representation is used to enhance the online interest group representation and the offline interest group representation, respectively:

[0034]

[0035]

[0036] wherein, denotes all interest groups and event nodes, G and I are the interest group set and the event set, respectively, denotes the edge set between the interest group and the published event, is the adjacency matrix, N and K represent the sizes of the event set and the interest group set, respectively, and wherein, is the interest group and event interaction matrix, when the interest group publishes the event , , denotes the interest group representation enhanced by events, denotes the trainable parameters of the graph neural network, is the graph neural network, and the calculation formula is wherein, denotes the i-th layer of the graph neural network ​​​​​​a feature embedding matrix of the layers, an adjacency matrix of a graph, a diagonal matrix a degree matrix of a node, an online representation of an interest group , an event-enhanced representation of an interest group , an offline representation of an interest group .

[0037] Further, the interest group recommendation method based on dual-intention contrastive learning in the event social network described above, wherein the step of aligning the online user representation, the offline user representation, the online enhanced interest group representation, and the offline enhanced interest group representation to obtain the dual-intention user representation and the dual-intention interest group representation of the fusion dual intention includes:

[0038] respectively constructing a contrastive loss at the user representation level and the enhanced interest group representation level to achieve cross-scene semantic fusion and complementary enhancement of the representation;

[0039] wherein the InfoNCE loss is used to realize the contrastive loss of the user representation, and is defined as follows:

[0040] ;

[0041] ;

[0042] The overall contrastive loss of the user representation is:

[0043] ;

[0044] The contrastive loss of the interest group representation is defined as follows:

[0045] ;

[0046] ;

[0047] The overall contrastive loss of the interest group representation is:

[0048] ;

[0049] wherein, c, respectively represent the user representation of the fusion of the online social intention and the offline participation intention, represents a similarity measurement function, U and G respectively are a user set and an interest group set, and respectively represent the interest group representation of the fusion of the online social intention and the offline participation intention.

[0050] Further, the interest group recommendation method based on double-intention contrastive learning in the event social network, wherein the expression of the double-intention user representation is:

[0051] ;

[0052] The expression of the double-intention interest group representation is:

[0053] ;

[0054] Wherein, c, respectively represent the user representation fused with online social and offline participation intention, and respectively represent the interest group representation fused with online social and offline participation intention, represents the cascading operation.

[0055] Further, the interest group recommendation method based on double-intention contrastive learning in the event social network, wherein the method further comprises:

[0056] Introducing a joint training framework of ranking loss and self-supervised loss when training the interest group recommendation:

[0057] ;

[0058] ;

[0059] ;

[0060] Wherein, D is a training data set, is a user participated interest group, is non-participated interest group, represents all trainable parameters in the framework, is a hyperparameter, is a hyperparameter for adjusting the contribution size of the contrastive loss, is the overall contrastive loss of the user representation, is the overall contrastive loss of the interest group representation, is the double-intention user representation, is the double-intention interest group representation.

[0061] Another object of the present application is to provide an interest group recommendation device based on double-intention contrastive learning in an event social network, comprising:

[0062] The first supergraph module is configured to construct an online interest group supergraph according to online interaction between a user and an interest group, perform representation learning on the online interest group supergraph through a supergraph neural network, and obtain online user representation and online interest group representation fused with a social intention;

[0063] The second supergraph module is configured to construct an offline event supergraph according to event data held by an interest group and member participation data, perform representation learning on the offline event supergraph through a supergraph neural network, and obtain offline user representation and offline interest group representation perceived with an offline intention;

[0064] The enhancement module is configured to enhance the online interest group representation and the offline interest group representation through interest group and event interaction, and obtain online enhanced interest group representation and offline enhanced interest group representation.

[0065] The alignment module is configured to align the online user representation, the offline user representation, the online enhanced interest group representation and the offline enhanced interest group representation, and obtain double-intention user representation and double-intention interest group representation fused with double intentions.

[0066] The recommendation module is configured to obtain a target user and an interest group, determine a target interest group according to a matching degree of the double-intention user representation and the double-intention interest group representation, and recommend the target interest group.

[0067] Another object of the present application is to provide a readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the steps of the above method.

[0068] Another object of the present application is to provide an electronic device comprising a memory, a processor and a computer program stored on the memory and running on the processor, the processor executing the program to implement the steps of the above method.

[0069] The present application constructs an online interest group hypergraph according to the online interaction relationship between the user and the interest group, performs representation learning on the online interest group hypergraph through a hypergraph neural network, and obtains online user representation and online interest group representation fused with social intent; constructs an offline event hypergraph according to event hosting by the interest group and member participation in the event data, performs representation learning on the offline event hypergraph through a hypergraph neural network, and obtains offline user representation and offline interest group representation with offline intent perception; respectively enhances the online interest group representation and the offline interest group representation through the interest group and event interaction relationship, and obtains online enhanced interest group representation and offline enhanced interest group representation; respectively aligns the online user representation, the offline user representation, the online enhanced interest group representation, and the offline enhanced interest group representation to obtain double-intent user representation and double-intent interest group representation fused with double intent; obtains a target user and an interest group, determines the target interest group according to the matching degree of the double-intent user representation and the double-intent interest group representation, and recommends the target interest group. The double-hypergraph structure is innovatively designed, that is, the online hyperedge is associated with the interest group and the member, and the offline hyperedge is connected with the interest group and the user and the event, the high-order dependent relationship between online and offline behaviors is simultaneously modeled through cross-hyperedge association of shared nodes, and the defect of one-sidedness of multi-dimensional behavior mode representation of the user is made up; in view of the limitation of single-intent modeling, based on the double-hypergraph structure, the user online social intent and offline event participation intent are learned respectively by using the hypergraph neural network, and a contrast learning mechanism is introduced to realize alignment and mutual enhancement of double-intent representation, so as to capture the dynamic change of the user interest, and then improve the accuracy of user demand recognition; in view of the static nature and semantic lack of the interest group feature representation, based on the interaction relationship between the interest group and the event, the GNN is used to aggregate the event information in the group, so that the interest group representation can accurately capture the difference and change of the event connotation, thereby more accurately matching the potential demand of the user. The problem of inaccurate interest group recommendation in the prior art is solved. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 The recommendation model network architecture diagram in the interest group recommendation method based on double-intent contrast learning in the event social network in the first embodiment of the present application;

[0071] Figure 2 The flowchart of the interest group recommendation method based on double-intent contrast learning in the event social network in the first embodiment of the present application;

[0072] Figure 3 The structural block diagram of the interest group recommendation device based on double-intent contrast learning in the event social network in the third embodiment of the present application.

[0073] The following specific embodiments will further illustrate the present application in conjunction with the above drawings. DETAILED DESCRIPTION

[0074] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0075] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0076] Unless otherwise defined, 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. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0077] Please see Figure 1 The diagram shows the network architecture of the recommendation model used in the interest group recommendation method based on dual-intent contrastive learning in an event social network proposed in an embodiment of the present invention. In this embodiment, a novel dual-intent-aware group recommendation model, GRICL, is proposed. Unlike existing research that only models online behavior, GRICL is the first to jointly model users' online social intent and offline participation intent within a unified framework. The designed dual-hypergraph structure achieves high-order dependency modeling of multiple entities online and offline through cross-hyperedge node associations, comprehensively depicting users' multi-dimensional behavioral patterns and providing a more realistic and comprehensive understanding of behavior for the recommendation system. Furthermore, unlike traditional contrastive learning which is only used for data augmentation or perspective consistency, the mechanism in this study conducts contrast at the intent level. By aligning heterogeneous behavioral intents in the online and offline hypergraphs, it achieves cross-scenario semantic fusion and mutual complementarity. This design can dynamically capture the evolution of user interests, thereby significantly improving the accuracy and adaptability of intent representation.

[0078] To address the issues of static and semantically incomplete feature representations of interest groups, a group-event bipartite graph is constructed. By using GNN to aggregate semantic information of events within a group, the group representation can dynamically reflect the differences and evolution of its activity content, thereby achieving a more accurate match for users' potential needs.

[0079] Specifically, it mainly includes four modules: Online Social Intent Learning, Offline Participation Intent Learning, Interest Group Representation Enhancement, Intent Alignment, Feature Alignment and Objective Optimization.

[0080] Among them, the Online Social Intent Learning module directly depicts the one-to-many relationship between online interest groups and members using hypergraph structure, that is, each hyperedge corresponds to an interest group and connects all member nodes of the group. Subsequently, with the help of the information propagation mechanism of hypergraph neural network, the deep aggregation and diffusion of social context information on the hypergraph are realized, and finally the refined user representation and interest group representation that integrate social intent are generated.

[0081] The Offline Participation Intent Learning module aims to mine the deep intent of users' offline participation events. To this end, the model first uses hypergraph to model the many-to-one interaction relationship between offline interest groups, events and users, thereby forming a double hypergraph structure with the online interest group hypergraph. By sharing user nodes in both hypergraphs, cross-scene high-order dependency relationships are captured. On this basis, the hypergraph neural network further extracts users' offline participation intent from event semantics and participation behavior, and learns user representations that integrate participation intent; finally, the offline view representation of the interest group is generated by aggregating user nodes.

[0082] The Interest Group Representation Enhancement module learns the interest group representation under the influence of events by constructing a bipartite graph structure between interest groups and events from the historical event interactions published by the interest group, and using graph neural network technology. Subsequently, the obtained interest group representation is integrated into the online interest group representation and offline interest group representation generated earlier, respectively, thereby giving these two aspects of representation more rich and comprehensive semantic information.

[0083] The Intent Alignment, Feature Alignment and Objective Optimization module introduces a contrastive learning mechanism based on intent hierarchy to align the semantics of online social intent and offline participation intent. Specifically, for the same user, the two kinds of intent representations corresponding to him in the double hypergraph are regarded as positive sample pairs, while the cross-scene intent representations of other users are regarded as negative sample pairs. By maximizing the similarity of positive samples and minimizing the similarity of negative samples, the consistency constraint of cross-scene intent is realized. Finally, combined with the recommendation optimization target based on the aligned user representation and interest group representation, the model minimizes the weighted sum of the recommendation loss and the intent alignment loss to realize the overall end-to-end optimization learning.

[0084] To further elaborate the specific implementation process of the embodiments of the present application, the meanings of some symbols, the definition of hypergraph representation and the interest group recommendation task are given as follows:

[0085] Specifically, given a user set , an event set , an interest group set , where is the size of the three sets. There are three types of interactions in the three sets , which are user-interest group interaction, user-event interaction and interest group-event interaction, and the corresponding interaction matrices are represented as , , respectively. The user-interest group interaction represents the historical joining of the interest group by the user, the user-event interaction represents the offline participation of the user in the event, and the interest group-event interaction represents the event publishing of the interest group.

[0086] As an extended graph structure, the unique feature of the hypergraph is that the hyperedge can be associated with any number of nodes. Formally, the hypergraph is defined as , where is the vertex set, is the hyperedge set, is the adjacency matrix of the hypergraph. If represents that the hyperedge connects the vertex , otherwise .

[0087] Given a user set U , an event set I , an interest group set G and their historical interaction data, the goal of the interest group recommendation task is to select several interest groups that the target user has not participated in and can best meet the needs of the target user from G as the recommended items.

[0088] For ease of discussion, Table 1 provides important symbols used in the embodiments of the present application.

[0089] Table 1

[0090]

[0091] The following will illustrate how to accurately recommend interest groups in the event social network by combining specific embodiments.

[0092] Embodiment One

[0093] Please refer to Figure 2The method based on double-intention contrast learning in the event social network in the first embodiment of the application is shown, and the method comprises steps S10-S14.

[0094] In step S10, an online interest group hypergraph is constructed according to an online interaction relationship between a user and an interest group, a hypergraph neural network is used for performing representation learning on the online interest group hypergraph, and an online user representation and an online interest group representation fused with a social intention are obtained.

[0095] First, three independent trainable embedding tables are constructed to store the vectorized representations of users, events and interest groups respectively. Since the model infers the user's inclination to the interest group through double intentions, it is assumed that each user and interest group contains both online social and offline participation features. Based on this assumption, an embedding decoupling mechanism is designed to decompose the original embedding space into two orthogonal subspaces, corresponding to the feature representation spaces of online and offline scenarios respectively. This fine-grained feature representation method enables the model to more effectively identify the differentiated behavior inclination of users in virtual socialization and entity activities, thereby obtaining better interest group matching capability.

[0096] In order to obtain the online and offline feature representations of users respectively, the following process is designed:

[0097] ;

[0098] represents a function of decomposing a vector into two sub-vectors, and the specific implementation can be a linear transformation function or a simple cutting operation. represents the online social representation of a user , represents the offline participation representation of a user . Accordingly, the online social representation matrix and the offline participation representation matrix of the user are and ;

[0099] Similarly, the online and offline feature embeddings of the interest group are obtained as follows:

[0100] ;

[0101] Let the online and offline feature embedding matrices of the interest group be , ;

[0102] In order to accurately describe the online interaction relationship between the user and the interest group, an online interest group hypergraph is constructed. The node set of the hypergraph not only contains all interest groups, but also covers all users, that is , is a superedge set, each interest group node is connected with its member nodes by a unique superedge, is the adjacency matrix of the online interest group supergraph. In particular, when a node is a member of an interest group , , the supergraph structure more completely preserves the multi-element interaction relationship between the interest group and the members than the ordinary graph structure.

[0103] On the basis of constructing the online interest group supergraph, the user representation and the online interest group representation fused with the social intention can be obtained by supergraph neural network representation learning. Specifically, the superedge is used as the medium for information transmission and processing in the supergraph neural network. First, the node information connected with the superedge is aggregated to generate a message; then the message is used to update the representation of each node associated with the superedge (i.e. the representation of the user or the interest group). After layer-by-layer propagation, the representation of the user and the interest group is finally improved from the rich superedge information, and the specific formula is as follows:

[0104] ;

[0105] wherein, , respectively represent the initial feature embedding matrix of the user, the interest group, is the supergraph neural network parameter. represents a classical supergraph neural network, and the calculation formula is wherein, represents the feature embedding matrix of the layer, represents the weight matrix of the layer, represents the adjacency matrix, and respectively represent the degree matrix of the node and the superedge.

[0106] Step S11, according to the event held by the interest group and the member participation event data, a offline event supergraph is constructed, and the offline user representation and the offline interest group representation of the offline intention perception are obtained by the supergraph neural network representation learning on the offline event supergraph.

[0107] ​Users often participate in offline events organized by online interest groups while joining them. Their intention to participate in offline events frequently influences their choice of online interest groups; therefore, this study focuses on user behavior patterns related to offline event participation. To uncover users' intentions for offline participation, a hypergraph structure is constructed based on user-event interaction data. A preference-aware hypergraph neural network is then employed to specifically characterize the complex relationships between users, events, and interest groups. This hypergraph neural network ultimately yields user representations and offline interest group representations that integrate offline participation intentions.

[0108] Specifically, a hypergraph is constructed based on data about events organized by interest groups and the participation of members. , in Represents a set of users and event nodes. Let each superedge be a set of superedges. Each interest group corresponds to a node in the interest group, and the nodes of the events they participate in are connected by a hyperedge. Represents the adjacency matrix of the offline event hypergraph, specifically, when nodes (in For interest groups Member set, for When the set of events in which members participate, .

[0109] Since each interest group is modeled as a hyperedge Therefore, it can be embedded from interest groups. Extracting super edges initial characterization Since a single hyperedge connects both member nodes and event nodes, and these two types of nodes have different semantic information, to maintain the distinguishability of these two types of nodes, message aggregation is divided into user message aggregation and event message aggregation, and user aggregated messages are obtained accordingly. Event aggregation messages The specific aggregation process is as follows:

[0110] ;

[0111] ;

[0112] in Indicates interest group The set of members, Indicates interest group members Embedded, Indicates interest group A set of events in which members participate. Indicates an event embedding of the super edge denotes a node aggregation function, which is generally a common attention aggregation mechanism.

[0113] aggregate two types of information to get the super edge enhanced representation. To avoid the aggregation process being dominated by active member information and weakening the contribution of events, an element-wise multiplication operation of event aggregation messages and interest group embeddings is introduced as a supplement to strengthen the contribution of the part of the event aggregation message that is highly related to the characteristics of the interest group in aggregation. Finally, the enhanced representation of the super edge The calculation formula is as follows:

[0114] ;

[0115] wherein denotes element-wise multiplication, denotes a trainable weight matrix, denotes a concatenation operation.

[0116] Finally, the updated super edge information is used to update the associated node representation. Formally, for a user node , the representation is updated as:

[0117] ;

[0118] wherein denotes the set of super edges associated with the user node , and denotes the embedding of the super edge associated with the user node .

[0119] To enhance the expression ability, a plurality of message aggregation propagation modules described above are further stacked to enable the node and super edge to represent the benefits from high-order neighbors. Finally, the average of the embeddings of each layer is taken to generate the final representation of the node, wherein the final representation of the user The calculation formula is as follows:

[0120] ;

[0121] wherein is the number of layers of the stacked message modules.

[0122] Since the super edge corresponds to an interest group, the final representation of the interest group involved in offline participation intention perception can be obtained by aggregating the embeddings of each layer of the super edge, and the formula is as follows:

[0123] .

[0124] ​​​Step S12, respectively enhance the online interest group representation and the offline interest group representation through the interest group-event interaction relationship, to obtain the online enhanced interest group representation and the offline enhanced interest group representation.

[0125] The interest group feature representation is enhanced through the interest group-event interaction relationship, specifically, the interest group-event relationship is represented as a bipartite graph , wherein represents all interest group and event nodes, represents a set of edges connected between the interest group and the published event, is an adjacency matrix and , wherein is an interest group-event interaction matrix, when the interest group publishes the event , .

[0126] Next, the concatenation of the interest group embedding and the event embedding , and the adjacency matrix are input into the graph neural network to refine the interest group representation in the following way:

[0127] ;

[0128] wherein, represents the representation of the interest group enhanced through the event, represents the trainable parameters of the graph neural network. is the graph neural network LightGCN, and its calculation formula is , wherein represents the feature embedding matrix of the layer, represents the adjacency matrix of the graph, the diagonal matrix represents the degree matrix of the node.

[0129] Finally, the initial embedding is convolved through L layers of convolutional layers, and the embedding obtained by each layer is averaged to obtain the final interest group representation enhanced by the event :

[0130] ;

[0131] The above representation is used to enhance the online and offline representations of the interest group respectively, and the formula is as follows:

[0132] ;

[0133] ;

[0134] wherein, online representation of interest group , event enhanced representation of interest group , offline representation of interest group .

[0135] Step S13, respectively aligning the online user representation, offline user representation, online enhanced interest group representation and offline enhanced interest group representation to obtain the fused dual-intention user representation and dual-intention interest group representation.

[0136] Among them, the contrast loss is constructed at the user representation level and the enhanced interest group representation level respectively to realize the fusion and complementary enhancement of cross-scene semantics;

[0137] Among them, the InfoNCE loss is used to realize the contrast loss of user representation, and is defined as follows:

[0138] ;

[0139] ;

[0140] The overall contrast loss of user representation is:

[0141] ;

[0142] The contrast loss of interest group representation is defined as follows:

[0143] ;

[0144] ;

[0145] The overall contrast loss of interest group representation is:

[0146] ;

[0147] Among them, c, respectively represent the user representation fused with online social and offline participation intentions, represents a similarity measurement function, U and G respectively represent the user set and the interest group set, and respectively represent the interest group representation fused with online social and offline participation intentions.

[0148] After obtaining the online and offline user representations respectively, they are combined to obtain the fused dual-intention user representation, and the final representation of the interest group is also obtained similarly, as follows:

[0149] The expression of the dual-intention user representation is:

[0150] ;

[0151] The expression of the dual-intention interest group representation is:

[0152] ;

[0153] Wherein, c, respectively represent the user representation of the fusion of online social and offline participation intention, and respectively represent the interest group representation of the fusion of online social and offline participation intention, represents the cascade operation.

[0154] Step S14, obtaining one of the target users and the interest group, determining the target interest group according to the matching degree determined by the dual-intention user representation and the dual-intention interest group representation, so as to recommend the target interest group.

[0155] Wherein, for the target user, the interest groups are sorted according to the matching degree (such as vector inner product similarity) of the dual-intention user representation and the dual-intention interest group representation, and the target interest group with high matching degree is recommended, that is, the target user from select several interest groups that have not participated in and can best meet the needs of the target user as the recommended items.

[0156] In summary, the interest group recommendation method based on double-intention contrast learning in the event social network in the above embodiments of the present application, by constructing an online interest group hypergraph according to the online interaction relationship between the user and the interest group, performing representation learning on the online interest group hypergraph through a hypergraph neural network to obtain online user representation and online interest group representation fused with social intention; constructing an offline event hypergraph according to the event held by the interest group and the event participation data of the members, performing representation learning on the offline event hypergraph through a hypergraph neural network to obtain offline user representation and offline interest group representation perceived with offline intention; respectively enhancing the online interest group representation and the offline interest group representation through the interest group and event interaction relationship to obtain online enhanced interest group representation and offline enhanced interest group representation; aligning the online user representation, the offline user representation, the online enhanced interest group representation, and the offline enhanced interest group representation to obtain double-intention user representation and double-intention interest group representation fused with double intention; obtaining a target user and an interest group, determining the target interest group according to the matching degree determined by the double-intention user representation and the double-intention interest group representation, and recommending the target interest group. The double-hypergraph structure is innovatively designed, that is, the online hyperedge is associated with the members and the offline hyperedge is connected with the interest group and the user and the event, the high-order dependency relationship between the online and offline behaviors is simultaneously modeled through the cross-hyperedge association of the shared nodes, and the defects of the one-dimensional behavior mode representation are made up; in view of the limitation of single-intention modeling, based on the double-hypergraph structure, the user online social intention and offline event participation intention are learned through a hypergraph neural network, and a contrast learning mechanism is introduced to realize the alignment and mutual enhancement of the double-intention representation, so as to capture the dynamic change of the user interest and further improve the accuracy of user demand recognition. In view of the static nature and semantic deficiency of the interest group feature representation, based on the interaction relationship between the interest group and the event, the GNN is used to aggregate the event information in the group, so that the interest group representation can accurately capture the difference and change of the event connotation, thereby more accurately matching the potential needs of the user. The problem of inaccurate interest group recommendation in the prior art is solved.

[0157] Embodiment two

[0158] The embodiment also proposes an interest group recommendation method based on double-intention contrast learning in an event social network. The difference between the interest group recommendation method based on double-intention contrast learning in the event social network in this embodiment and the interest group recommendation method based on double-intention contrast learning in the event social network in embodiment one is that:

[0159] The method further comprises:

[0160] In the training of the interest group recommendation, a joint training framework of ranking loss and self-supervised loss is introduced:

[0161] ;

[0162] ;

[0163] ;

[0164] where, is the training data set, is the interest group that the user participates in, is the interest group that the user does not participate in, denotes all trainable parameters in this framework, is the hyper-parameter, is the hyper-parameter that adjusts the contribution size of the contrastive loss, is the overall contrastive loss that characterizes the user, is the overall contrastive loss that characterizes the interest group, is the dual-intention user representation, is the dual-intention interest group representation.

[0165] where, for a target user and an interest group , the predicted score of the user to the interest group is calculated using the inner product, and the calculation formula is as follows:

[0166] ;

[0167] Then, the Bayesian personalized ranking (BPR) loss is used for optimization, as shown below:

[0168] ;

[0169] where, is the training data set, is the interest group that the user participates in, is the interest group that the user does not participate in, denotes all trainable parameters in this framework, where is the hyper-parameter,

[0170] Finally, the ranking loss and the self-supervised loss are unified into a joint training framework:

[0171] ;

[0172] where, is the hyper-parameter that adjusts the contribution size of the contrastive loss.

[0173] In summary, the interest group recommendation method based on double-intention contrast learning in the event social network in the above embodiments of the present application, by constructing an online interest group hypergraph according to the online interaction relationship between the user and the interest group, performing representation learning on the online interest group hypergraph through a hypergraph neural network to obtain online user representation and online interest group representation fused with social intention; constructing an offline event hypergraph according to event data held by the interest group and event participation data of members, performing representation learning on the offline event hypergraph through a hypergraph neural network to obtain offline user representation and offline interest group representation perceived with offline intention; respectively enhancing the online interest group representation and the offline interest group representation through the interest group and event interaction relationship to obtain online enhanced interest group representation and offline enhanced interest group representation; respectively aligning the online user representation, the offline user representation, the online enhanced interest group representation and the offline enhanced interest group representation to obtain double-intention user representation and double-intention interest group representation fused with double intention; obtaining a target user and an interest group, determining the target interest group according to the matching degree of the double-intention user representation and the double-intention interest group representation to recommend the target interest group. The double-hypergraph structure is innovatively designed, that is, the online hyperedge is associated with the interest group and the member, and the offline hyperedge is connected with the interest group and the user and the event, the high-order dependency relationship between online and offline behaviors is simultaneously modeled through cross-hyperedge association of shared nodes, and the defect of one-sidedness of user multi-dimensional behavior mode representation is made up; in view of the limitation of single-intention modeling, based on the double-hypergraph structure, the user online social intention and offline event participation intention are learned respectively by using the hypergraph neural network, and a contrast learning mechanism is introduced to realize alignment and mutual enhancement of double-intention representation, so as to capture the dynamic change of user interest and further improve the accuracy of user demand recognition. In view of the static nature and semantic deficiency of the interest group feature representation, based on the interaction relationship between the interest group and the event, the GNN is used to aggregate the event information in the group, so that the interest group representation can accurately capture the difference and change of the event connotation, thereby more accurately matching the potential needs of the user. The problem of inaccurate interest group recommendation in the prior art is solved.

[0174] Embodiment three

[0175] Please refer to Figure 3 , which is an interest group recommendation device based on double-intention contrast learning in an event social network according to a third embodiment of the present application, the device comprises:

[0176] The first hypergraph module 100 is used to construct an online interest group hypergraph according to the online interaction relationship between the user and the interest group, perform representation learning on the online interest group hypergraph through a hypergraph neural network to obtain online user representation and online interest group representation fused with social intention;

[0177] The second supergraph module 200 is configured to construct an offline event supergraph according to event-organizing interest group data and member-participating event data, and perform representation learning on the offline event supergraph by using a supergraph neural network to obtain offline user representation and offline interest group representation with offline intent perception.

[0178] The enhancement module 300 is configured to enhance the online interest group representation and the offline interest group representation respectively by using interest group and event interaction relationship to obtain online enhanced interest group representation and offline enhanced interest group representation.

[0179] The alignment module 400 is configured to obtain fused double-intent user representation and double-intent interest group representation by respectively aligning the online user representation, the offline user representation, the online enhanced interest group representation and the offline enhanced interest group representation.

[0180] The recommendation module 500 is configured to obtain a target user and an interest group, determine a target interest group according to the matching degree of the double-intent user representation and the double-intent interest group representation, and recommend the target interest group.

[0181] The functions or operation steps realized by the above modules when executed are basically the same as those of the above method embodiments, and thus will not be described here.

[0182] Embodiment Four

[0183] In another aspect, the present application also provides a readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to realize the steps of the method according to any one of the above embodiments.

[0184] Embodiment Five

[0185] In another aspect, the present application also provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to realize the steps of the method according to any one of the above embodiments.

[0186] The technical features of each of the above embodiments can be combined in any manner, and to make the description concise, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present application.

[0187] Those skilled in the art will appreciate that the logic and / or steps represented in the flow diagrams, or otherwise described herein, can be embodied in

[0188] More specific examples (a non-exhaustive list) of the computer readable storage medium include the following: an electrical connection having one or more wires (electrical device), a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer readable storage medium can also be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example, via optical scanning of the paper or other medium, then compiled, interpreted, or otherwise processed in a suitable manner, if necessary, and then stored in a computer memory.

[0189] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the above embodiments, various steps or methods can be implemented in software or firmware that is stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any of the following technologies, known in the art, or combinations thereof, can be used: a discrete logic circuit having logic gates for implementing logic functions upon data signals, an application specific integrated circuit having appropriate combinational logic gates, a programmable gate array (PGA), a field programmable gate array (FPGA), or the like.

[0190] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Also, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in an appropriate manner.

[0191] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the patent scope of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for interest group recommendation based on dual intent contrastive learning in an event social network, characterized in that, The method comprises: According to the online interaction relationship between the user and the interest group, an online interest group hypergraph is constructed, representation learning of the online interest group hypergraph is performed through a hypergraph neural network, online user representation and online interest group representation fused with a social intention are obtained, and specifically: The expression of the online interest group hypergraph is: ； Fusing social intent with online user representation and online interest group representation is expressed as: ； wherein, , U and G are the set of users, the set of interest groups, respectively, is the set of hyper-edges, is the adjacency matrix of the online interest group hypergraph, M , K denote the size of the set of users and the set of interest groups, respectively, when the node wherein is the set of members of the interest group , , , denote the initial feature embedding matrices of users, interest groups, respectively, are the hypergraph neural network parameters, denotes a classical hypergraph neural network; According to the event held by the interest group and the event participation data of the members, an offline event hypergraph is constructed, representation learning of the offline event hypergraph is performed through a hypergraph neural network, offline user representation and offline interest group representation perceived with an offline intention are obtained, and specifically: The expression of the offline event hypergraph is: ； Offline user representation and offline interest group representation The expression is: ； ； in, Represents a set of users and event nodes. Let each superedge be a set of superedges. Corresponding to an interest group, Interest group member nodes and the event nodes they participate in are connected via hyperedges. This represents the adjacency matrix of the offline event hypergraph, where M, N, and K represent the sizes of the user set, event set, and interest group set, respectively. When the node... hour, ,in, For interest groups Member set, for A set of events in which members participate. The number of layers in which the message modules are stacked. l This is the layer index for the message module. , Represents user node The set of associated hyperedges Represents user node Associated hyperedge Embedded, ,in This indicates element-wise multiplication. This represents a trainable weight matrix. This indicates a splicing operation. , , Indicates interest group The set of members, Indicates interest group members Embedded, Indicates an event Embedded, This represents the node aggregation function, which is an attention aggregation mechanism. d This indicates the dimension of the embedding; Online interest group representation and offline interest group representation are respectively enhanced through the interaction relationship between the interest group and the event, online enhanced interest group representation and offline enhanced interest group representation are obtained, and specifically: Expressing interest group event interaction as a bipartite graph ; Embedding interest groups with events and adjacency matrices into a graph neural network, resulting in: ； The initial embedding is passed through L layers of convolutional layers, and the embedding obtained by each layer is averaged to obtain a final event-lifted interest group representation : ； Characterizing interest groups Respectively for enhancing online interest group characterization and offline interest group characterization: ； ； wherein, represents all interest group and event nodes, G and I are interest group set and event set respectively, represents the set of edges connecting interest groups and published events, is an adjacency matrix, N, K represent the size of event set and interest group set respectively, and wherein, is the interest group and event interaction matrix, when the interest group publishes an event , , represents the representation of interest group enhanced by events, represents the trainable parameters of graph neural network, is a graph neural network, represents the online representation of interest group , represents the event promotion representation of interest group , represents the offline representation of interest group ; After alignment of the online user representation, the offline user representation, the online enhanced interest group representation and the offline enhanced interest group representation, double-intention user representation and double-intention interest group representation fused with double intentions are obtained, and specifically: Contrast loss is constructed on the user representation level and the enhanced interest group representation level to realize fusion and complementary enhancement of cross-scene semantics; One of the target users and the interest group is obtained, the matching degree determined according to the double-intention user representation and the double-intention interest group representation is used to determine the target interest group, and the target interest group is recommended.

2. The method of claim 1, wherein, The InfoNCE loss is used to realize the contrast loss of the user representation, and is defined as follows: ； ； The overall contrast loss of the user representation is: ； The contrast loss of the interest group representation is defined as follows: ； ； The overall contrast loss of the interest group representation is: ； wherein, c, respectively represent user representations that fuse online social and offline participation intentions, represents a similarity measure function, U and G respectively are a set of users, a set of interest groups, and respectively represent interest group representations that fuse online social and offline participation intentions.

3. The method of claim 2, wherein the method is based on a dual intent contrastive learning in an event social network. The expression of the double-intention user representation is: ； The expression of the double-intention interest group representation is: ； wherein, c, respectively represent user representations that fuse online social and offline engagement intent, and respectively represent interest group representations that fuse online social and offline engagement intent, represents a cascading operation.

4. The method of claim 3, wherein, The method further comprises: A joint training framework of ranking loss and self-supervised loss is introduced when training the interest group recommendation: ； ； ； wherein, is the training data set, is the user participating interest group, is the non-participating interest group, denotes all trainable parameters in the framework, is the hyperparameter, is the hyperparameter that adjusts the contribution size of the contrastive loss, is the overall contrastive loss represented by the user, is the overall contrastive loss represented by the interest group, is the dual-intent user representation, is the dual-intent interest group representation.

5. An apparatus for interest group recommendation based on dual intent contrastive learning in event social networks, characterized in that, The device is used to realize the interest group recommendation method based on double-intention contrast learning in the event social network, and the device comprises: A first hypergraph module is configured to construct an online interest group hypergraph according to the online interaction relationship between the user and the interest group, perform representation learning of the online interest group hypergraph through a hypergraph neural network, and obtain online user representation and online interest group representation fused with a social intention; A second hypergraph module is configured to construct an offline event hypergraph according to the event held by the interest group and the event participation data of the members, perform representation learning of the offline event hypergraph through a hypergraph neural network, and obtain offline user representation and offline interest group representation perceived with an offline intention; An enhancement module is configured to enhance online interest group representation and offline interest group representation through the interaction relationship between the interest group and the event, and obtain online enhanced interest group representation and offline enhanced interest group representation; An alignment module is configured to obtain double-intention user representation and double-intention interest group representation fused with double intentions after alignment of online user representation, offline user representation, online enhanced interest group representation and offline enhanced interest group representation; A recommendation module is configured to obtain one of the target users and the interest group, determine the target interest group according to the matching degree determined according to the double-intention user representation and the double-intention interest group representation, and recommend the target interest group.

6. A readable storage medium, having stored thereon a computer program, characterized in that, The program, when executed by the processor, implements the steps of the method as claimed in any one of claims 1 to 4.

7. An electronic device, comprising: A computer program product comprising a memory, a processor, and a computer program stored on the memory and running on the processor, the processor implementing the steps of the method as claimed in any one of claims 1 to 4 when executing the program.

Images