A community-aware graph-enhanced interest point recommendation method

Abstract

This invention discloses an interest point recommendation method based on community-aware graph enhancement, comprising the following steps: obtaining user node embeddings and interest point node embeddings on the user-interest graph using a graph neural network; discovering user communities and interest point communities using a user community classifier and an interest point community classifier, respectively; discovering community relationships between user communities and interest point communities using a community relationship encoder; reconstructing user-interest graph training parameters based on user communities, interest point communities, and community relationships to generate a community graph; and generating node-enhanced graphs and edge-enhanced graphs using intra-community similarity and inter-community connectivity, respectively; constructing user multi-graph embeddings and interest point multi-graph embeddings and training them based on Bayesian personalized ranking loss; given a query user, calculating their rating of candidate interest points and recommending interest points to the user. This invention can improve the performance of interest point recommendation and has broad application prospects in advertising, information push, and other fields.

Description

Technical Field

[0001] This invention belongs to the field of information recommendation technology, specifically relating to an interest point recommendation method based on community-aware graph enhancement. Background Technology

[0002] With the rapid development of smartphones and other smart devices equipped with GPS sensors, location-based social networking services (LBSNs) have emerged, such as Foursquare, Facebook Places, and Google Places. On LBSNs, users can check in and share at points of interest, such as shops and restaurants. LBSNs have a large user base and cover a wide geographical area. By mining check-in data to recommend points of interest, they can not only help users discover new areas of interest and explore unfamiliar regions, but also facilitate advertisers in pushing ads to target users.

[0003] Given that graph structures can effectively model the complex relationships between users and points of interest (POIs), graph-based recommendation methods transform historical check-in data between users and POIs into a graph and utilize graph neural networks (GNNs) to obtain the embeddings of users and POIs. Despite their effectiveness, existing graph-based methods can suffer from poor performance due to missing and noisy check-ins.

[0004] Chinese patent application CN116738051A discloses an interest point recommendation method based on graph neural networks and collaborative relationship mining. This method constructs a bipartite graph of user-interest point relationships, using users and interest points as nodes and the check-in relationships between users and interest points as edges. Based on the implicit relationship nodes of user nodes in the graph and the user's check-in time for interest points, it extracts the feature representations of all hidden layers of the user nodes in the graph neural network using a multi-head self-attention mechanism. A softmax algorithm is then used to calculate the interest point recommendation list, recommending interests with a probability value higher than a set threshold to the user. This scheme focuses on mining the implicit attributes of nodes and may not be effective for handling complex relationships and higher-order interaction patterns.

[0005] Chinese patent application CN116805020A discloses an interest point recommendation method based on graph neural networks and context-aware information. It extracts high-order nonlinear interactions between nodes through a graph structure and models spatiotemporal and social influences. The constructed graph is convolved to output user and interest point representations. These representations are then weighted to obtain the final user and interest point representations, and an inner product operation is used to generate the final predicted value. This scheme utilizes auxiliary information such as user check-in time, location, and interest point category to alleviate data sparsity. However, the model is relatively complex, and changes in the recommendation scenario may affect its generalization ability, leading to a decline in recommendation quality.

[0006] Graph augmentation addresses noise and sparsity by generating modified graph data, while edge augmentation serves as the optimal augmentation strategy. Heuristic edge augmentation methods sample edges uniformly or based on weights computed according to the graph structure. Parametric edge augmentation methods use downstream tasks to guide edge addition and removal: some methods directly parameterize the adjacency matrix; some add and remove edges based on node labels predicted by deep neural networks; and some first compute connection probabilities between nodes based on trainable embeddings and then sample edges accordingly. The first two parametric methods are designed for classification tasks and cannot be applied to recommendation. While heuristic and the last type of parametric methods can be applied to recommendation, they still have some drawbacks: heuristic methods cannot perform adaptive recommendation augmentation; and parametric methods ignore user-user and interest-point-interest correlations when adding and removing edges between users and points of interest. Summary of the Invention

[0007] In view of the above, the purpose of this invention is to provide an interest point recommendation method based on community-aware graph enhancement. This method groups users with similar preferences and interest points with similar characteristics into user communities and interest point communities, and performs graph enhancement by reconstructing user-interest point graphs through relationships between communities. Edges are added and deleted according to community affiliation and relationships between communities, thereby improving the recommendation performance of interest points. This method is suitable for application scenarios such as advertising and information push.

[0008] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0009] This invention provides a method for recommending points of interest based on community-aware graph enhancement, comprising the following steps:

[0010] User node embeddings and interest point node embeddings are obtained on the user-interest graph using graph neural networks. Based on the user node embeddings and interest point node embeddings, user communities and interest point communities are discovered using user community classifiers and interest point community classifiers, respectively. Community relationship encoders are used to discover community relationships between user communities and interest point communities.

[0011] The training parameters of the user-interest graph are reconstructed based on user communities, interest point communities, and community relationships to generate community graphs. Node-enhanced graphs and edge-enhanced graphs are generated by utilizing intra-community similarity and inter-community connectivity, respectively.

[0012] The user multi-graph embedding and interest point multi-graph embedding are obtained by combining the embeddings of the reconstructed user-interest point graph, node augmentation graph, edge augmentation graph and community graph. The user rating of the interest point is calculated, and the user multi-graph embedding and interest point multi-graph embedding are trained based on Bayesian personalized ranking loss.

[0013] Given a query user, the system calculates the query user's rating of candidate interest points based on the trained user multi-graph embedding and interest point multi-graph embedding, and recommends interest points to the user according to the rating ranking.

[0014] Preferably, obtaining user node embeddings and interest point node embeddings on the user-interest graph using a graph neural network includes:

[0015] Input User-Point of Interest Map Initialize user node embeddings and interest point node embeddings. Use the lightweight graph convolutional network LightGCN and take the initialized user node embeddings and interest point node embeddings as input to the first layer of LightGCN. Then, obtain the updated user node embeddings and interest point node embeddings v through a total of L layers of graph convolution.

[0016] In LightGCN, the graph convolution operation at layer l+1 is defined as follows:

[0017]

[0018]

[0019] In the formula, u i and v j Let i and j represent the i-th user and the j-th point of interest, respectively. and Let N(·) represent the output of the embedding of the i-th user node and the output of the embedding of the j-th interest point node in the (l+1)-th layer, respectively, and let N(·) represent the one-hop neighbors of the node. and Let represent the inputs of the i-th user node embedding and the j-th interest point node embedding in the (l+1)-th layer, respectively, where 1 ≤ l ≤ L.

[0020] Preferably, the discovery of user communities and point-of-interest (POI) communities based on user node embedding and point-of-interest (POI) node embedding, using user community classifiers and POI community classifiers respectively, includes:

[0021] Given the user node embedding u, following the idea of ​​user-based collaborative filtering, we treat the set of similar users as user communities, use a user community classifier to obtain the probability of each user belonging to each user community, and build a user community affiliation matrix Z. U , means as follows:

[0022]

[0023] In the formula, Z represents the user community affiliation matrix. U The i-th row, u i Let represent the embedding of the i-th user node, mlip denotes the function implemented by the multilayer perceptron, and the softmax function ensures that the sum of all probabilities is 1;

[0024] Given an input interest point node embedding v, following the idea of ​​project-based collaborative filtering, a set of similar interest points is considered as an interest point community. An interest point community classifier is used to obtain the probability that each interest point belongs to each interest point community and to build an interest point affiliation matrix Z. I , means as follows:

[0025]

[0026] In the formula, Z represents the community affiliation matrix of points of interest. I The j-th line, v j This indicates the embedding of the j-th interest point node.

[0027] Preferably, the step of discovering community relationships between user communities and point-of-interest communities using a community relationship encoder includes:

[0028] Input user community affiliation matrix Z U Interest Point Community Affiliation Matrix Z I The user community and interest point community are obtained based on the embedding of lower-level users and interests, respectively, and are represented as follows: and T represents the transpose. Based on these embeddings, the community relationship matrix B between user communities and interest-based communities is obtained using a community encoder, as shown below:

[0029]

[0030] In the formula, B mn p represents the m-th user community m And the nth interest point community q n The relationship between community relations is the element at the position of the m-th row and n-th column of the community relation matrix B. The sigmoid function ensures that the value of the community relation is between 0 and 1.

[0031] Preferably, the step of reconstructing user-interest graph training parameters based on user communities, interest-point communities, and community relationships to generate a community graph includes:

[0032] First, input the user community affiliation matrix Z. U Interest Point Community Affiliation Matrix Z I Given the community relationship matrix B, the user-interest graph reconstructor is used to establish connections between users and interests in connected user communities and interest communities to reconstruct the user-interest graph, as shown below:

[0033]

[0034] In the formula, Indicates user u i and points of interest v j The probability that an edge exists between them;

[0035] Secondly, based on the reconstruction loss of the user-interest graph, the parameters of user node embedding u and interest point node embedding v, as well as the user community classifier, interest point community classifier, and community relationship encoder, are trained, outputting the trained user community affiliation matrix, interest point community affiliation matrix, and community relationship matrix. The reconstruction loss L is then used. r It is expressed as follows:

[0036]

[0037] In the formula, e ij =1 or 0, indicating user u i and points of interest v j Does an edge actually exist between them?

[0038] Furthermore, a sparse user community affiliation matrix is ​​obtained by sparsifying the trained user community affiliation matrix and interest point community affiliation matrix. and interest-based community affiliation matrix

[0039] Finally, the sparse community relation matrix B is obtained by sparsifying the trained community relation matrix. S Thus, a community graph is constructed.

[0040] Preferably, the step of generating node-enhanced graphs and edge-enhanced graphs using intra-community similarity and inter-community connectivity respectively includes:

[0041] Using a sparse user community affiliation matrix and interest-based community affiliation matrix To identify relationships within a community, based on the similarity between users or points of interest within the same community, users or points of interest on the edges are replaced with other similar users or points of interest within the same user community or point of interest community to fill in missing check-ins, thereby generating a node augmentation graph.

[0042] Using a sparse user community affiliation matrix Interest Point Community Affiliation Matrix Community Relationship Matrix B S To identify relationships between communities, based on the connectivity between communities, edges are added between users and points of interest connected to the same community to fill in missing check-ins, and edges are removed between users and points of interest not connected to the same community to eliminate noisy check-ins, thus generating an edge-enhanced graph.

[0043] Preferably, the embedding of the reconstructed user-interest point graph, node augmentation graph, edge augmentation graph, and community graph yields user multi-graph embedding and interest point multi-graph embedding, and the calculation of user ratings for interest points includes:

[0044] First, input the node augmentation graph. User node augmentation embedding u is obtained using graph neural networks and graph convolution. n Enhanced embedding of interest point nodes v n Input edge augmentation graph User edge augmentation embedding u is obtained by using graph neural networks and graph convolution. e And interest point edge augmentation embedding v e Input community map Embedded p in the initial user community 0 Embedding with the initial interest community q 0 The embeddings p of user communities and q of interest point communities are obtained using graph neural networks and graph convolution. 0 and q 0 It is expressed as follows:

[0045]

[0046]

[0047] Secondly, based on the user community affiliation matrix Z U Interest Point Community Affiliation Matrix Z I The user's community-based embedding *p* and the interest-point community embedding *q* are aggregated to obtain the user's community-based embedding *u*. c And interest-based community-based embedding v c , means as follows:

[0048]

[0049]

[0050] In the formula, This represents the community-based embedding of the i-th user. This represents the community-based embedding of the j-th point of interest;

[0051] Then, concatenate the user node embedding u and the user node enhanced embedding u. n User-side enhancement embedding u e And user-based community-embedded u c Get user multi-image embedding u f ; Concatenate interest point node embedding v, interest point node enhanced embedding v n Interest point edge enhancement embedding v e And interest-based community-embedded v c , obtain interest point multi-graph embedding v f ;

[0052] Embedding u based on user multiple images f and point of interest multi-image embedding v f Get user u i For point of interest v j rating It is expressed as follows:

[0053]

[0054] In the formula, This represents the multi-image embedding for the i-th user. This represents the multi-graph embedding of the j-th point of interest.

[0055] Preferably, the step of training user multi-graph embeddings and interest point multi-graph embeddings based on Bayesian personalized ranking loss includes:

[0056] The prediction loss is defined as the Bayesian personalized ranking loss, expressed as follows:

[0057]

[0058] In the formula, v j Indicates user u i Points of interest that have been checked in, v j′ Indicates user u i Points of interest obtained from sampling have never been checked in;

[0059] By minimizing the prediction loss, the user node embedding u and the interest point node embedding v are sufficiently optimized, and the optimized user multi-graph embedding u is obtained accordingly. f And optimized point-of-interest multi-graph embedding v f .

[0060] Preferably, for a given query user, the query user's rating of candidate interest points is calculated based on the trained user multi-graph embedding and interest point multi-graph embedding, and interest points are recommended to the user according to the rating ranking, including:

[0061] Given a query user u i The score for each candidate point of interest is calculated and represented as follows:

[0062]

[0063] In the formula, Indicates query user u i For point of interest v j The rating, This represents the i-th optimized user multi-graph embedding. This represents the optimized multi-graph embedding of the j-th point of interest;

[0064] The calculated scores are sorted in descending order, and the top k points of interest are returned as recommendations.

[0065] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0066] (1) The method of the present invention uses a community classifier and a community relationship encoder to discover user communities and interest communities as well as the relationships between communities, and reconstructs the user-interest graph, which can be used for graph enhancement and can achieve adaptive recommendation enhancement.

[0067] (2) The method of the present invention generates node enhancement graph and edge enhancement graph based on the similarity within the community and the connectivity between communities, respectively. Edges are added and deleted according to the communities that can reflect user preferences and interest point characteristics. The method fully considers the correlation between users and interest points, as well as between users and interest points and between interest points, and further improves the interest point recommendation performance. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0069] Figure 1 This is a flowchart illustrating the interest point recommendation method based on community-aware graph enhancement provided in an embodiment of the present invention.

[0070] Figure 2 This is a logical framework diagram of the interest point recommendation method based on community-aware graph enhancement provided in the embodiments of the present invention. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this invention clearer, the invention 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 merely illustrative of the invention and do not limit the scope of protection of this invention.

[0072] The inventive concept of this invention is as follows: In order to solve the technical problems of sparsity and noise in graph-based point of interest recommendation, such as how to add missing edges related to point of interest recommendation and delete noisy edges unrelated to point of interest recommendation, this invention provides a community-aware graph-enhanced point of interest recommendation method. By grouping users with similar preferences and points of interest with similar characteristics into user communities and point of interest communities, edges are added and deleted according to community affiliation and relationships between communities, thereby further improving the performance of point of interest recommendation.

[0073] Figure 1 This is a framework diagram of the interest point recommendation method based on community-aware graph enhancement provided in an embodiment of the present invention. Figure 1 As shown in the example, this embodiment provides an interest point recommendation method based on community-aware graph enhancement, which generally includes the following steps:

[0074] S1. On the user-interest graph, the user node embedding and interest point node embedding are obtained by using a graph neural network. Based on the user node embedding and interest point node embedding, the user community classifier and the interest point community classifier are used to discover the user community and the interest point community, respectively. The community relationship encoder is used to discover the community relationship between the user community and the interest point community.

[0075] S2, based on user communities, interest point communities and community relationships, reconstruct user-interest point graph training parameters to generate community graphs, and use intra-community similarity and inter-community connectivity to generate node-enhanced graphs and edge-enhanced graphs respectively;

[0076] S3. Combine the embeddings of the reconstructed user-interest point graph, node augmentation graph, edge augmentation graph and community graph to obtain user multi-graph embedding and interest point multi-graph embedding. Calculate the user's rating of the interest point and train the user multi-graph embedding and interest point multi-graph embedding based on Bayesian personalized ranking loss.

[0077] S4, Given a query user, calculate the query user's rating of candidate interest points based on the trained user multi-graph embedding and interest point multi-graph embedding, and recommend interest points to the user according to the rating ranking.

[0078] Figure 2 This is a logical framework diagram of the interest point recommendation method based on community-aware graph enhancement provided in this embodiment of the invention. Figure 2As shown in the embodiment, an interest point recommendation method based on community-aware graph enhancement is provided, which specifically includes the following steps:

[0079] Step 1, Input User-Point of Interest Map Initialize the node embeddings for users and points of interest, and update them using GNN.

[0080] In typical recommendation scenarios, users and points of interest usually only have their IDs as input, without semantic meaning. Multiple nonlinear transformations, including feature transformations and activation functions, do not help learn better embeddings. On the contrary, they may increase the difficulty of training the model. Therefore, we utilize a GNN that only includes neighbor aggregation—LightGCN—to update the embeddings of users and points of interest. The (l+1)th layer graph convolution operation in LightGCN is defined as follows:

[0081]

[0082]

[0083] In the formula, u i and v j Let i and j represent the i-th user and the j-th point of interest, respectively. and Let N(·) represent the output of the embedding of the i-th user node and the output of the embedding of the j-th interest point node in the (l+1)-th layer, respectively, and let N(·) represent the one-hop neighbors of the node. and Let represent the inputs of the i-th user node embedding and the j-th interest point node embedding in the (l+1)-th layer, respectively, where 1 ≤ l ≤ L.

[0084] Input User-Point of Interest Map The initialized user and interest point embeddings are used as input to the first layer of LightGCN, and the updated user node embedding u and interest point node embedding v are obtained through L-layer graph convolution.

[0085] Step 2: Input user node embedding u and interest point node embedding v, and obtain the user community affiliation matrix Z using the user community classifier and interest point community classifier respectively. U Interest Point Community Affiliation Matrix Z I .

[0086] Collaborative filtering is one of the most popular recommendation methods. User-based collaborative filtering and interest-based collaborative filtering are two typical methods. The former aims to find similar users, while the latter aims to find similar interests. The set of similar users and the set of similar interests can be considered as user communities and interest communities, respectively. Users within the same community typically have similar preferences; interests within the same community typically have similar characteristics.

[0087] To discover user communities and interest-based communities from user behavior data through learning, user community classifiers and interest-based community classifiers are used to obtain user u i and points of interest v j The probability of belonging to each community is expressed as follows:

[0088]

[0089]

[0090] In the formula, Z represents the user community affiliation matrix. U The i-th row, u i This indicates the embedding of the i-th user node. Z represents the community affiliation matrix of points of interest. I The j-th line, v j Let represent the embedding of the j-th interest point node, and let mlip denote the function implemented by the multilayer perceptron. The softmax function ensures that the sum of all probabilities is 1.

[0091] Step 3, input the user community affiliation matrix Z U Interest Point Community Affiliation Matrix Z I The community relationship matrix B between user communities and interest point communities is obtained by using a community encoder.

[0092] The transpose of the user community affiliation matrix and the point of interest (POI) community affiliation matrix can be viewed as the embeddings of user communities and POI communities based on their respective subordinate users and POIs, denoted as follows: and Based on these embeddings, the user community p is obtained using the community encoder. m and interest point community q n Relationship B mn , means as follows:

[0093]

[0094] In the formula, B mn p represents the m-th user community m And the nth interest point community q n The relationship between community relations is the element at the position of the m-th row and n-th column of the community relation matrix B. The sigmoid function ensures that the value of the community relation is between 0 and 1.

[0095] Step 4, input the user community affiliation matrix Z U Interest Point Community Affiliation Matrix Z I Based on the community relationship matrix B, the user-interest graph is reconstructed using a user-interest graph reconstructor.

[0096] Without appropriate supervision signals, it is difficult to learn the relationship between user communities and point-of-interest (POI) communities. Here, graph reconstruction is introduced. The core idea is that users and POIs in connected user communities and POI communities are also connected to reconstruct the user-POI graph, which is consistent with the idea of ​​the random block model, as shown below:

[0097]

[0098] in, Indicates user u i and points of interest v j The probability that an edge exists between them.

[0099] Step 5, based on user-interest graph The reconstruction loss is used to train the embeddings u and v of users and points of interest, as well as the parameters of the user community classifier, point of interest community classifier, and community relationship encoder, and outputs the trained user community affiliation matrix, point of interest community affiliation matrix, and community relationship matrix.

[0100] By minimizing the user-interest graph The reconstruction loss, user and interest point embeddings u and v, user community classifier, interest point community classifier, and community relationship encoder parameters can be fully optimized to obtain a suitable community affiliation matrix and community relationship matrix. Reconstruction Loss L r Specifically, it is expressed as follows:

[0101]

[0102] Among them, e ij =1 or 0, indicating user u i and points of interest v j Does an edge actually exist between them?

[0103] Step 6: Input the community relationship matrix and obtain a sparse community relationship matrix B through sparsification. S Thus, a community graph is constructed.

[0104] By inputting a community relationship matrix, a community graph can be constructed through sparsification. Among them, if the user community p m and interest point community q n Relationship B mn If the value is greater than or equal to the edge threshold ∈, then There exists an edge e in it. mn ,otherwise There is no corresponding edge in B; if B mn If ≥∈, then the sparse community relation matrix B S The element at the position of the m-th row and n-th column otherwise

[0105] Step 7, input the user community affiliation matrix Z U Interest Point Community Affiliation Matrix Z I A sparse user community affiliation matrix is ​​obtained through sparsification. and interest-based community affiliation matrix

[0106] By assigning users and points of interest to the corresponding communities with the highest probabilities, the trained user community affiliation matrix and point of interest community affiliation matrix are sparsified to obtain a sparse user community affiliation matrix. and interest-based community affiliation matrix Among them, if but otherwise if but otherwise Where m and n represent the serial numbers of the user community and the interest point community, respectively.

[0107] Step 8, Input a sparse user community affiliation matrix and interest-based community affiliation matrix Enhanced graph by considering node similarity within the community.

[0108] Considering the similarity between nodes within the same community, replacing edge users or points of interest with users or points of interest from the same user community or point of interest community can help fill in missing check-ins. A sparse community affiliation matrix is ​​used to identify relationships within communities, and a node augmentation graph is generated based on this matrix.

[0109] First, Node copy Then, for Each edge e in ij If the user replacement probability is satisfied Then from u i A user is randomly selected from the community. s Replace u i ,generate edge e in sj Otherwise, if the probability of interest point replacement is satisfied... Then from v j Randomly select a point of interest v in the community t Replace v j ,generate edge e in it Otherwise, generate directly. edge e in ij .

[0110] Step 9, Input a sparse user community affiliation matrix Interest Point Community Affiliation Matrix Community Relationship Matrix B S By considering the connectivity between communities, an edge-enhanced graph is generated.

[0111] Considering that nodes in connected communities are more likely to connect than nodes in unconnected communities, adding edges between connected users and points of interest within the same community can help fill in missing check-ins, while removing edges between unconnected users and points of interest within the same community can help eliminate noisy check-ins. A sparse community affiliation matrix and a sparse community relationship matrix are used to identify relationships between communities, and an edge-enhanced graph is generated based on these.

[0112] First, Copy the nodes and edges to Then, for each user-interest pair (u i ,v j ): 1) If u i and v j exist The middle is not connected, while u i and v j The community it belongs to If the middle is connected, then the probability is added to the edge. Towards Add edge e ij ;2) If u i and v j exist The middle phase is connected, and u i and v j The community it belongs to If the edges are not connected, then the probability of edge deletion is used. from Delete edge e ij .

[0113] Step 10, Input the node augmentation graph Using GNN to obtain user node augmented embeddings u n Enhanced embedding of interest point nodes v n .

[0114] Input node augmented graph User node embedding u and interest point node embedding v are used as inputs to the first layer of LightGCN, and the enhanced user node embedding u is obtained through L-layer graph convolution. n Enhanced embedding of interest point nodes v n .

[0115] Step 11, Input the edge enhancement graph Using GNN to obtain user-side augmented embeddings u e And interest point edge augmentation embedding v e .

[0116] Input edge augmentation graph User node embedding u and interest point node embedding v are used as inputs to the first layer of LightGCN, and the user edge augmented embedding u is obtained through L-layer graph convolution. e And interest point edge augmentation embedding v e .

[0117] Step 12, Input Community Map Use GNN to obtain user community embedding p and interest point community embedding q.

[0118] First, the embedding p of the initial user community is calculated based on the community affiliation relationship. 0 Embedding with the initial interest community q 0 , means as follows:

[0119]

[0120]

[0121] Enter community map Embedding the initial user community p 0 Embedding with the initial interest community q 0 As input to the first layer of LightGCN, the embedding p of the user community and the embedding q of the interest point community are obtained through L-layer graph convolution.

[0122] Step 13, input the user community affiliation matrix Z U Interest Point Community Affiliation Matrix Z I And the embedding p of the user community and the embedding q of the point-of-interest community, calculate the user's community-based embedding u. c And interest-based community-based embedding v c .

[0123] Aggregating user community embeddings and interest-based community embeddings according to community affiliation yields the community-based embeddings of users and interests, as shown below:

[0124]

[0125]

[0126] In the formula, This represents the community-based embedding of the i-th user. This represents the community-based embedding of the j-th point of interest.

[0127] Step 14: Concatenate the user node embedding and the point of interest node embedding u and v, and the user node augmentation embedding and the point of interest node augmentation embedding u. n and v n User-enhanced embedding and interest-based enhanced embedding u e and v e And user-based community embedding and interest-based community embedding u c and v c To obtain user multi-image embedding and point-of-interest multi-image embedding u f and v f Based on this, the system calculates the user's rating of the points of interest.

[0128] Augmented graphs and community graphs can be used to model users and points of interest across different views. Combining embeddings from different views can better model user check-ins at points of interest. i and points of interest v j The multi-graph embedding is obtained by concatenating the embeddings of nodes, edges, and communities, as shown below:

[0129]

[0130]

[0131] Embedding u based on user multiple images f and point of interest multi-image embedding v f , obtain u i For v j rating It is expressed as follows:

[0132]

[0133] In the formula, This represents the multi-image embedding for the i-th user. This represents the multi-graph embedding of the j-th point of interest.

[0134] Step 15: Train user node embeddings and interest point node embeddings u and v based on Bayesian personalized ranking loss to obtain the optimized user multi-graph embedding u. f And optimized point-of-interest multi-graph embedding v f .

[0135] The prediction loss is defined as the Bayesian personalized ranking loss, expressed as follows:

[0136]

[0137] In the formula, v j Indicates user u iPoints of interest that have been checked in, v j′ Indicates user u i Points of interest obtained from sampling have never been checked in.

[0138] By minimizing the prediction loss, the user node embedding u and the interest point node embedding v are sufficiently optimized, and the optimized user multi-graph embedding u is obtained accordingly. f And optimized point-of-interest multi-graph embedding v f .

[0139] Step 16, given the query user u i It calculates the score for each candidate point of interest and returns the top k points of interest as the recommendation result.

[0140] Given a query user u i The score for each candidate point of interest is calculated and represented as follows:

[0141]

[0142] In the formula, Indicates query user u i For point of interest v j The rating, This represents the i-th optimized user multi-graph embedding. This represents the optimized multi-graph embedding of the j-th point of interest;

[0143] The calculated scores are sorted in descending order, and the top k points of interest are returned as recommendations.

[0144] In summary, the interest point recommendation method based on community-aware graph enhancement provided by this invention utilizes a collaborative filtering strategy, a community classifier, and a community relationship encoder to discover user communities, interest point communities, and relationships between communities. It reconstructs a user-interest graph for graph enhancement, and simultaneously generates node-enhanced graphs and edge-enhanced graphs based on intra-community similarity and inter-community connectivity. Edges are added and deleted according to communities that reflect user preferences and interest point characteristics, enabling adaptive recommendation enhancement. Furthermore, it fully considers the correlations between users and interest points, as well as between users and interest points, further improving interest point recommendation performance. It can also mine user preferences from a vast user check-in history and recommend potentially interesting interest points, improving recommendation performance and avoiding the significant time and effort required for manual search. This method has broad application prospects in areas such as advertising and information push.

[0145] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for recommending points of interest based on community-aware graph enhancement, characterized in that, Includes the following steps: User node embeddings and interest point node embeddings are obtained from the user-interest graph using a graph neural network. Based on user node embedding and point-of-interest (POI) node embedding, user community classifiers and POI community classifiers are used to discover user communities and POI communities respectively, including: input user node embedding Following the idea of ​​user-based collaborative filtering, this approach treats sets of similar users as user communities, uses a user community classifier to obtain the probability of each user belonging to each user community, and constructs a user community affiliation matrix. , means as follows: ， In the formula, Represents the user community affiliation matrix The OK, Indicates the first Embedded user nodes This represents a function implemented by a multilayer perceptron. The function ensures that the sum of all probabilities is 1; input interest point node embedding. Following the idea of ​​project-based collaborative filtering, this approach treats sets of similar points of interest as interest communities, uses an interest community classifier to obtain the probability that each point of interest belongs to each interest community, and then constructs an interest community affiliation matrix. , means as follows: ， In the formula, Represents the community affiliation matrix of points of interest The OK, Indicates the first Embedding of interest point nodes; Discovering community relationships between user communities and interest-based communities using a community relationship encoder includes: inputting a user community affiliation matrix. and interest-based community affiliation matrix The user community and interest point community are obtained based on the embedding of lower-level users and interests, respectively, and are represented as follows: and , The representation is transposed. Based on these embeddings, a community relationship matrix between user communities and interest-based communities is obtained using a community encoder. , means as follows: ， In the formula, Indicates the first User Community and the Interest Points Community The relationships between them, i.e., the community relationship matrix The Line 1 The element at the position of the column, The function ensures that the community relationship value ranges between 0 and 1; Reconstruct user-interest graph training parameters based on user communities, interest-point communities, and community relationships to generate community graphs; Node-enhanced graphs and edge-enhanced graphs are generated by leveraging intra-community similarity and inter-community connectivity, including: using a sparse user community affiliation matrix. and interest-based community affiliation matrix To identify relationships within a community, based on the similarity between users or points of interest within the same community, users or points of interest on the edges are replaced with other similar users or points of interest within the same user community or point of interest community to fill in missing check-ins, thereby generating a node augmentation graph. Use a sparse user community affiliation matrix Interest Point Community Affiliation Matrix Community Relationship Matrix To identify relationships between communities, based on the connectivity between communities, edges are added between users and points of interest connected to the same community to fill in missing check-ins, and edges are removed between users and points of interest not connected to the same community to eliminate noisy check-ins, thus generating an edge-enhanced graph. ; The user multi-graph embedding and interest point multi-graph embedding are obtained by combining the embeddings of the reconstructed user-interest point graph, node augmentation graph, edge augmentation graph and community graph. The user rating of the interest point is calculated, and the user multi-graph embedding and interest point multi-graph embedding are trained based on Bayesian personalized ranking loss. Given a query user, the system calculates the query user's rating of candidate interest points based on the trained user multi-graph embedding and interest point multi-graph embedding, and recommends interest points to the user according to the rating ranking.

2. The interest point recommendation method based on community-aware graph enhancement according to claim 1, characterized in that, The step of obtaining user node embeddings and interest point node embeddings using a graph neural network on the user-interest graph includes: Input User-Point of Interest Map Initialize user node embeddings and interest point node embeddings. Utilize the lightweight graph convolutional network LightGCN, using the initialized user node embeddings and interest point node embeddings as input to the first layer of LightGCN. Through... Layered graph convolution obtains updated user node embeddings embedding of interest point nodes ; Among them, the first in LightGCN The layer graph convolution operation is defined as follows: ， ， In the formula, and They represent the first The user and the first One point of interest and They represent the first The first layer The output embedded by the user node and the first The output of embedding each point of interest node This represents a node's one-hop neighbor. and They represent the first The first layer The input embedded in the first user node and the first The input is embedded with a point of interest node. .

3. The interest point recommendation method based on community-aware graph enhancement according to claim 1, characterized in that, The process of reconstructing user-interest graph training parameters based on user communities, interest-point communities, and community relationships to generate a community graph includes: First, input the user community affiliation matrix. Interest Point Community Affiliation Matrix Community Relationship Matrix The user-interest graph reconstructor is used to establish connections between users and points of interest in connected user communities and interest communities to reconstruct the user-interest graph, as shown below: ， In the formula, Indicates user and points of interest The probability that an edge exists between them; Secondly, user node embeddings are trained based on the reconstruction loss of the user-interest graph. embedding of interest point nodes The algorithm includes parameters for the user community classifier, interest point community classifier, and community relationship encoder, outputting the trained user community affiliation matrix, interest point community affiliation matrix, and community relationship matrix, and reconstructing the loss function. It is expressed as follows: ， In the formula, Or 0, indicating the user and points of interest Does an edge actually exist between them? Furthermore, a sparse user community affiliation matrix is ​​obtained by sparsifying the trained user community affiliation matrix and interest point community affiliation matrix. and interest-based community affiliation matrix ; Finally, a sparse community relation matrix is ​​obtained by sparsifying the trained community relation matrix. Thus, a community graph is constructed. .

4. The interest point recommendation method based on community-aware graph enhancement according to claim 1, characterized in that, The embedding of the reconstructed user-interest graph, node augmentation graph, edge augmentation graph, and community graph yields user multi-graph embedding and interest point multi-graph embedding. User ratings for interest points are then calculated, including: First, input the node augmentation graph. User node augmentation embeddings are obtained using graph neural networks via graph convolution. Enhanced embedding of interest point nodes Input edge augmentation graph User edge augmentation embeddings are obtained using graph neural networks via graph convolution. Interest Point Edge Augmentation Embedding Input community map Embedding the initial user community Embedding with the initial interest community The embeddings of user communities are obtained by using graph neural networks and graph convolution. Embedding with interest-based communities , and It is expressed as follows: ， ； Secondly, based on the user community affiliation matrix and interest-based community affiliation matrix Embedding in user communities Embedding with interest-based communities Aggregation yields user community-based embeddings Community-based embedding of points of interest , means as follows: ， ， In the formula, Indicates the first Individual user-based community embedding, Indicates the first Each point of interest is embedded in the community; Then, concatenate the user node embedding. Enhanced embedding of user nodes User-side enhancement embedding and user-based community embedding Get user multi-image embedding ; Embedding of point of interest nodes Interest Point Node Enhancement Embedding Interest Point Edge Enhancement Embedding And interest-based community embedding Obtaining multi-image embedding of interest points ; Based on user-embedded multiple images Multi-image embedding with points of interest Get users Points of interest rating , means as follows: ， In the formula, Indicates the first Multiple image embedding per user, Indicates the first Multiple graph embedding of points of interest.

5. The interest point recommendation method based on community-aware graph enhancement according to claim 4, characterized in that, The training of user multi-graph embeddings and interest point multi-graph embeddings based on Bayesian personalized ranking loss includes: The prediction loss is defined as the Bayesian personalized ranking loss, expressed as follows: ， In the formula, Indicates user Points of interest that have been checked in Indicates user Points of interest obtained from sampling have never been checked in; By minimizing the prediction loss, user node embedding embedding of interest point nodes The optimization is fully achieved, resulting in an optimized user multi-image embedding. And optimized point-of-interest multi-graph embedding .

6. The interest point recommendation method based on community-aware graph enhancement according to claim 5, characterized in that, For a given query user, based on the trained user multi-graph embedding and interest point multi-graph embedding, the query user's rating of candidate interest points is calculated, and interest points are recommended to the user according to the rating ranking, including: Given query user The score for each candidate point of interest is calculated and represented as follows: ， In the formula, Indicates the query user Points of interest The rating, Indicates the first An optimized user multi-image embedding Indicates the first Optimized multi-image embedding of points of interest; Sort the calculated scores in descending order and return the top-ranked scores. The points of interest are used as the recommendation results.

Images