Method for recommending next interest point of temporary group based on hybrid graph neural network

Abstract

The invention relates to the field of group prediction recommendation, in particular to a method for recommending a next interest point of a temporary group based on a hybrid graph neural network, which comprises the following steps of: fusing spatio-temporal information of group members and feature information of the members to generate fitting representation of the temporary group, learning fitting feature representation containing overall features and dynamic preferences of the temporary group members, and recommending the next interest point of the temporary group. The problem brought by dynamic preference generated by temporary group members due to temporal and spatial changes is solved, and the accuracy of recommending the next interest point to the temporary group is improved; the popularity difference between the interest points is reflected through directed edges and weights thereof, interaction preference representation of members to the interest points is obtained through further learning, and the influence of the popularity of the interest points on preference changes of the interest points of the members is fully considered; an interest point transfer directed graph is constructed based on an access sequence of members to interest points, transfer preference representation of the members to the interest points is obtained through learning, and context information is fully considered while a potential place conversion trend of a user is considered.

Description

Technical Field

[0001] The present invention relates to the technical field of group prediction recommendation, and in particular to a method for recommending the next point of interest of a temporary group based on a hybrid graph neural network. Background Art

[0002] As one of the important services of location-based social networks, the next point of interest (POI) recommendation provides users with personalized location recommendations by analyzing their geographic location, behavioral habits, and social connections, enhancing the social network experience and bringing more commercial value to service providers. Among them, the next POI recommendation for groups is a research hotspot. By mining the check-in records and other available information of group members, the next POI that group members may visit at a specific time and place is predicted.

[0003] At present, existing research mainly focuses on POI recommendation for fixed groups, and there are fewer studies on POI recommendation for temporary groups. Temporary groups are composed of strangers or acquaintances who gather at a specific time and place. These members usually lack long-term communication and interaction records, resulting in large differences in preferences among members, which has become a major difficulty in POI group recommendation in current research. Although there are studies that use hybrid graph neural networks to recommend POI lists for temporary groups, they ignore the dynamic changes of member preferences with time and space information and cannot meet the immediate needs of members.

[0004] In addition, in the field of next POI recommendation in groups, some studies have proposed the use of graph convolutional networks and attention mechanism technology to extract the connection between group and individual preferences, but this study did not consider the differences between the current spatiotemporal information of group members and the contextual information between POIs.

[0005] In order to improve the accuracy of POI recommendation, it is necessary to fully consider the spatiotemporal state of the group in forming the group fitting representation and consider the contextual information between POIs in the POI transition graph to selectively emphasize which more relevant historical check-ins are the model that best meets the needs of members at the moment for the next POI selection of the temporary group. Summary of the invention

[0006] The present invention provides a method for recommending the next point of interest of a temporary group based on a hybrid graph neural network.

[0007] The technical solution of the present invention is as follows:

[0008] The method for recommending the next point of interest in a temporary group based on a hybrid graph neural network includes the following steps:

[0009] S1. Temporary group fitting feature representation: obtain the social relationships of members in the temporary group, build a social network graph, and obtain the social influence of each member in the temporary group based on the degree of nodes, the shortest path length between nodes, and the length of edges in the social network graph;

[0010] Based on the node's temporal feature vector, spatial feature vector, temporal weight matrix and spatial weight matrix, the node's spatiotemporal attention weight is obtained; based on the spatiotemporal attention weight, the learnable weight matrix, the vector displacement and the transposed vector, the comprehensive attention coefficient is obtained, and the neighbor nodes are aggregated and multi-headed attention operations are performed in sequence based on the comprehensive attention coefficient to generate the node's spatiotemporal information representation;

[0011] Based on the obtained social influence and spatiotemporal information representation, a weighted sum is performed to obtain the fitting feature representation of the temporary group;

[0012] S2. Learning the interaction preferences of members’ interest points: Based on the interaction records between members and interest points, we construct the connection edges between members, between members and interest points, and between interest points and interest points. Based on the popularity of interest points, we get the popularity weights of the connection edges between interest points and interest points, and build a bipartite graph of the interaction between members and interest points.

[0013] Based on popularity weights and learnable weight matrices, we can obtain information about the spread of members and POIs, and between POIs.

[0014] Based on the popularity weight, learnable parameter vector, transposed vector, learnable weight matrix, and spatiotemporal attention weight, the comprehensive popularity weight coefficient is obtained. The neighbor messages are further aggregated based on the comprehensive popularity weight coefficient to obtain the aggregated member messages and point of interest messages.

[0015] Perform multi-layer mapping processing on the aggregated member messages and POI messages, and update the member's interactive preference representation for POIs;

[0016] S3. Learning members’ interest point transfer preferences: Based on the members’ visit order to interest points, a directed graph of interest point transfer is constructed, with interest points as nodes and visit relationships between interest points as edges.

[0017] Based on the normalized weights of the connecting edges, the features of the interest points themselves, the weight matrix of the features of the interest points themselves, the features with context information transmitted from the neighboring interest points, the weight matrix of the features with context information transmitted from the neighboring interest points, and the bias vector, after weighted summation processing, the transfer preference representation of the members for the interest points is obtained;

[0018] S4. Recommend the next point of interest: sum the members' interaction preference representations for the points of interest and the members' transfer preference representations for the points of interest to obtain a final point of interest representation, perform vector product operations and normalization operations on the final point of interest representation and the fitting feature representation of the temporary group, obtain the score of the temporary group on each candidate point of interest, and determine the order of recommending the points of interest to the temporary group based on the scores.

[0019] Specifically, in S3, the normalized weight of the connecting edge, the feature of the point of interest itself, the weight matrix of the feature of the point of interest itself, the feature with context information transmitted from the neighboring point of interest, the weight matrix of the feature with context information transmitted from the neighboring point of interest, and the bias vector are weighted and summed to obtain the member's transfer preference representation for the point of interest. Specifically:

[0020]

[0021] in, Point of interest p i Combined with its neighboring interest points p j Characteristics at time t and its own characteristics x i , A ij From the point of interest p i To the point of interest j is the normalized weight of the connecting edge, P is the weight matrix with context information features passed from the neighboring interest points, Q is the weight matrix of the interest point’s own features, and b represents the bias vector.

[0022] Specifically, in S2, based on the popularity weight and the learnable weight matrix, the message propagation between members and points of interest, and between points of interest, is obtained, which is specifically expressed as:

[0023]

[0024] in, and The message is transferred from the interest point p based on the positive edge and the negative edge of the lth layer respectively. j Propagate to member u i , w pj Point of interest p j is the popularity weight, MSG is the message propagation function, is the member u at level l i The expression, is the interest point p at the lth layer j The expression, For the lth layer based on the positive edge from the interest point p j To member u i The spread of news, For the lth layer based on negative edges from the interest point p j To member u i the spread of news;

[0025]

[0026] in, and The lth layer is based on positive edges and negative edges to send messages from member u to j Propagate to the point of interest p i , w pi Point of interest p i is the popularity weight, MSG is the message propagation function, is the member u at level l j The expression, is the interest point p at the lth layer i The expression, For the lth layer based on the positive edge from member u j To the point of interest i The spread of news, For the lth layer based on the negative edge from member u j To the point of interest i the spread of news;

[0027]

[0028] in, and The message is transferred from the interest point p based on the positive edge and the negative edge of the lth layer respectively. j Propagate to the point of interest p i , w pj Point of interest p j is the popularity weight, MSG is the message propagation function, is the interest point p at the lth layer i The expression, is the interest point p at the lth layer j The expression, For the lth layer based on the positive edge from the interest point p j To the point of interest i The spread of news, For the lth layer based on negative edges from the interest point p j To the point of interest i of news spread.

[0029] Furthermore, in S2, a comprehensive popularity weight coefficient is obtained based on the popularity weight, the learnable parameter vector, the transposed vector, the learnable weight matrix, and the spatiotemporal attention weight, and neighbor message aggregation is further performed based on the comprehensive popularity weight coefficient to obtain aggregated member messages and point of interest messages, respectively, which are specifically expressed as follows:

[0030]

[0031] Among them, α' ij is the comprehensive popularity weighted coefficient, W∈R F'*F is a learnable weight matrix, F is the dimension of the input feature, F' is the dimension of the output feature, || is a cascade operation, is the learnable parameter vector, LeakyReLU(·) is the activation matrix, N i For member u i The neighbor set of For member u i and member u j The spatiotemporal attention coefficient, w pj and w pk They are points of interest p j and points of interest k The popularity weight of , exp(·) is the Softmax function used to normalize the weight coefficient;

[0032]

[0033] Among them, N i is the neighbor set of node i, α' ij is the comprehensive popularity weighted coefficient, for node u i , whose message m l (u i ) is through u i For neighbor node u j Message l (u j ) is obtained by weighted summation; for node p i , whose message m l (p i ) is through p i For neighbor node p j Message l (p j ) is obtained by weighted summation.

[0034] Furthermore, in S2, multi-layer mapping processing is performed on the aggregated member messages and the POI messages, and the member's interaction preference representation for the POI is updated, which is specifically expressed as:

[0035]

[0036] in, is the representation of the member set, is the representation of the interest point set, MLP is a neural network model consisting of two fully connected neural layers stacked with loss rate and activation mechanism, w pj Point of interest p j The popularity weight, w pi Point of interest p i The popularity weight of is the representation of the updated l+1th layer member set, is the representation of the updated l+1th layer interest point set, and For each member u i Four sets of information obtained from neighbors, and For each point of interest p i Four sets of information obtained from neighbors.

[0037] Specifically, in S1, the spatiotemporal attention weight of the node is obtained based on the temporal feature vector, spatial feature vector, temporal weight matrix and spatial weight matrix of the node. The formula is:

[0038]

[0039] in, For node u i and u j The spatiotemporal attention weight, α is a hyperparameter used to balance the spatial and temporal relationship, which is a value between 0 and 1; LeakyReLU(·) is the activation function, and are the time weight matrix and the space weight matrix respectively, and The node u i and u j The time feature vector of and The node u i and u j Spatial eigenvector, i and j are the serial numbers of the nodes.

[0040] Furthermore, in S1, the comprehensive attention coefficient α is obtained based on the spatiotemporal attention weight, the learnable weight matrix, the vector displacement and the transposed vector. ij , specifically expressed as:

[0041]

[0042] in, and For node u i and node u j Representation, W∈R F'*F is a learnable weight matrix, F is the dimension of the input feature, F' is the dimension of the output feature, || represents a cascade operation, for , R is the social network radius of the member, LeakyReLU(·) is the activation matrix, N i For node u i Neighbors, For node u i and u j The spatiotemporal attention coefficient of , exp(·) is the normalization coefficient of the Softmax function, and T is the transposed vector.

[0043] Furthermore, in S1, based on the degree of nodes in the social network graph, the shortest path length between nodes, and the length of edges, the social influence of each member in the temporary group is obtained, and the formula is expressed as:

[0044]

[0045] Where INF(i) is social influence, R is the social network radius of the member, and R is set to 1; N x (i) is node u i x th The order of neighbors, d ij For node u i and node u j The shortest path length between j For node u j The degree of node u i The nth neighbor of ij is the weight of the social relationship between members, and ω is set ij =1.

[0046] Furthermore, the normalized weight of the connecting edge in S3 is expressed as follows:

[0047]

[0048] in, is the normalized weight of the incoming edge connection, is the normalized weight of the outgoing edge connection, count(v i ,v j )∈{0,1}, count(v j ,v i ) both represent node v i and node vj Is there an edge between them? If yes, the value is 1; if no, the value is 0; count(v i ,v k )∈{0,1}, count(v k ,v i ) both represent node v i and node v k Whether there is an edge between them, if yes, the value is 1, if not, the value is 0; For node v i Predecessor node set, For node v i Post node set.

[0049] A temporary group next point of interest recommendation system based on hybrid graph neural network, characterized by comprising:

[0050] Temporary group fitting feature representation module: used to obtain the social relationships of members in the temporary group, build a social network graph, and obtain the social influence of each member in the temporary group based on the degree of nodes, the shortest path length between nodes, and the length of edges in the social network graph; used to obtain the spatiotemporal attention weight of the node based on the temporal feature vector, spatial feature vector, temporal weight matrix, and spatial weight matrix of the node; based on the spatiotemporal attention weight, learnable weight matrix, vector displacement, and transposed vector, obtain the comprehensive attention coefficient, and generate the spatiotemporal information representation of the node after performing aggregation and multi-head attention operations on neighbor nodes in turn based on the comprehensive attention coefficient; used to obtain the fitting feature representation of the temporary group after weighted summation based on the obtained social influence and spatiotemporal information representation;

[0051] Learning member interest point interaction preference module: based on the interaction records between members and interest points, construct connection edges between members and members, between members and interest points, and between interest points and interest points, obtain the popularity weights of the connection edges between interest points and interest points based on the popularity of interest points, construct an interactive bipartite graph of members and interest points, obtain the message propagation between members and interest points and interest points and interest points based on the popularity weights and the learnable weight matrix, obtain the comprehensive popularity weighting coefficient based on the popularity weights, learnable parameter vectors, transposed vectors, learnable weight matrices, and spatiotemporal attention weights, further aggregate neighbor messages based on the comprehensive popularity weighting coefficient, obtain aggregated member messages and interest point messages, respectively, perform multi-layer mapping processing on the aggregated member messages and interest point messages, and update to obtain the member's interaction preference representation for interest points;

[0052] Learning member interest point transfer preference module: used to construct an interest point transfer directed graph with interest points as nodes and access relations between interest points as edges based on the member's visit order to interest points.

[0053] Based on the normalized weights of the connecting edges, the features of the interest points themselves, the weight matrix of the features of the interest points themselves, the features with context information transmitted from the neighboring interest points, the weight matrix of the features with context information transmitted from the neighboring interest points, and the bias vector, after weighted summation processing, the transfer preference representation of the members for the interest points is obtained;

[0054] Recommendation module: used to obtain the final interest point representation by summing up the members' interaction preference representations for interest points and the members' transfer preference representations for interest points, and to obtain the score of the temporary group on each candidate interest point after performing vector product operation and normalization operation on the final interest point representation and the fitting feature representation of the temporary group, and to determine the order of recommending interest points to the temporary group based on the score.

[0055] The present invention also provides a temporary group next point of interest recommendation system based on hybrid graph neural network, comprising:

[0056] Temporary group fitting feature representation module: used to obtain the social relationships of members in the temporary group, build a social network graph, and obtain the social influence of each member in the temporary group based on the degree of nodes, the shortest path length between nodes, and the length of edges in the social network graph; used to obtain the spatiotemporal attention weight of the node based on the temporal feature vector, spatial feature vector, temporal weight matrix, and spatial weight matrix of the node; based on the spatiotemporal attention weight, learnable weight matrix, vector displacement, and transposed vector, obtain the comprehensive attention coefficient, and generate the spatiotemporal information representation of the node after performing aggregation and multi-head attention operations on neighbor nodes in turn based on the comprehensive attention coefficient; used to obtain the fitting feature representation of the temporary group after weighted summation based on the obtained social influence and spatiotemporal information representation;

[0057] Learning member interest point interaction preference module: based on the interaction records between members and interest points, construct connection edges between members and members, between members and interest points, and between interest points and interest points, obtain the popularity weights of the connection edges between interest points and interest points based on the popularity of interest points, construct an interactive bipartite graph of members and interest points, obtain the message propagation between members and interest points and interest points and interest points based on the popularity weights and the learnable weight matrix, obtain the comprehensive popularity weighting coefficient based on the popularity weights, learnable parameter vectors, transposed vectors, learnable weight matrices, and spatiotemporal attention weights, further aggregate neighbor messages based on the comprehensive popularity weighting coefficient, obtain aggregated member messages and interest point messages, respectively, perform multi-layer mapping processing on the aggregated member messages and interest point messages, and update to obtain the member's interaction preference representation for interest points;

[0058] Learning member interest point transfer preference module: used to construct an interest point transfer directed graph with interest points as nodes and access relations between interest points as edges based on the member's visit order to interest points.

[0059] Based on the normalized weights of the connecting edges, the features of the interest points themselves, the weight matrix of the features of the interest points themselves, the features with context information transmitted from the neighboring interest points, the weight matrix of the features with context information transmitted from the neighboring interest points, and the bias vector, after weighted summation processing, the transfer preference representation of the members for the interest points is obtained;

[0060] Recommendation module: used to obtain the final interest point representation by summing up the members' interaction preference representations for interest points and the members' transfer preference representations for interest points, and to obtain the score of the temporary group on each candidate interest point after performing vector product operation and normalization operation on the final interest point representation and the fitting feature representation of the temporary group, and to determine the order of recommending interest points to the temporary group based on the score.

[0061] The beneficial effects of the present invention are:

[0062] 1. The present invention is oriented to temporary groups and takes into account spatiotemporal context information. It fuses the spatiotemporal information of group members with their feature information to generate a fitting representation of the temporary group to form a virtual user. It learns a fitting feature representation containing the overall features and dynamic preferences of temporary group members, solves the problem caused by the dynamic preferences of temporary group members due to spatiotemporal changes, and improves the accuracy of recommending the next point of interest for the temporary group.

[0063] 2. The present invention uses the popularity of interest points to enhance the learning members' interest point interaction preferences, constructs an interactive bipartite graph of members and interest points, reflects the popularity differences between interest points through directed edges and their weights, and further learns to obtain the members' interactive preference representation for interest points, fully considering the impact of interest point popularity on the change of members' interest point preferences;

[0064] 3. The present invention captures the complex transfer relationship between points of interest, constructs a directed graph of point of interest transfer based on the order in which members visit points of interest, learns members' transfer preference representations for points of interest, and fully considers contextual information while considering users' potential location conversion trends. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In the attached picture:

[0066] Figure 1 It is a schematic diagram of the next point of interest recommendation process for a temporary group based on a hybrid graph neural network in an embodiment;

[0067] Figure 2The following is a comparison of the recommendation effects of the STC-NPROG model under different group sizes on the Fouesquare, Gowalla, and Yelp datasets in the embodiments. DETAILED DESCRIPTION

[0068] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings.

[0069] Example

[0070] This embodiment provides a method for recommending the next point of interest in a temporary group based on a hybrid graph neural network. Figure 1 , including the following steps:

[0071] S1: Temporary group fitting feature representation: obtain the social relationships of members in the temporary group, build a social network graph, and obtain the social influence of each member in the temporary group based on the degree of nodes, the shortest path length between nodes, and the length of edges in the social network graph;

[0072] Based on the node's temporal feature vector, spatial feature vector, temporal weight matrix and spatial weight matrix, the node's spatiotemporal attention weight is obtained; based on the spatiotemporal attention weight, the learnable weight matrix, the vector displacement and the transposed vector, the comprehensive attention coefficient is obtained, and the neighbor nodes are aggregated and multi-headed attention operations are performed in sequence based on the comprehensive attention coefficient to generate the node's spatiotemporal information representation;

[0073] Based on the obtained social influence and spatiotemporal information representation, a weighted sum is performed to obtain the fitting feature representation of the temporary group.

[0074] In the present invention, it is first necessary to clearly define the relevant issues and concepts, as follows:

[0075] User set: Let U = {u1,u2,…,u m} represents the user set, where m represents the number of users. Each user in U is distinguished by a unique user id.

[0076] POI set: A POI is defined as a geographical location with a specific function that can meet certain user needs (such as a movie theater or a cafe). In this study, POI contains three attributes: a unique identifier id, semantic information c, and geographic information d (lon, lat), where lon represents longitude and lat represents latitude. In this study, let P = {p1, p2, …, p n} represents the POI set, and n represents the number of POIs.

[0077] Temporary group: A temporary group is a group of users (belonging to U) who gather together for an activity. The group will disband after the activity ends. Members of a temporary group have different social influences due to their social backgrounds and rights. In this work, a temporary group is defined as OG = {u i ,u j ,…,u r}, r represents the size of OG, and OG is a subset of U.

[0078] Social Network: Let G s =<V,E> Represents the social network of members of the temporary group OG, where V = {u1,u2,…,u m} represents G S Node, E represents G S The edge of ij = i ,u j >(i≠j,e ij ∈GS) represents the member u i and u j There are social relationships between them.

[0079] POI check-in sequence set: Let S = {S i ,S j ,…,S r} represents the set of POI check-in sequences of all members in OG, where S i ={s1,s2,…,s i} represents member u i The sign-in sequence set, and s i =p1,p2,…,p k}, k∈[0,n] represents member u i The i-th POI check-in sequence.

[0080] Spatiotemporal information: Each member u i ∈U corresponds to a spatiotemporal information e in OG i , space-time information i It consists of the member's check-in history and time information, represented by e i =(loc i , time i ), where loc i Represents geographic location, time i Represents a timestamp. The information of members at different times and locations affects their behavior and social influence in temporary groups.

[0081] ​Based on the above definition, the next POI recommendation problem facing a temporary group in the present invention can be described as: based on the above given relevant data, recommend a next POI that satisfies all group members to the temporary group OG.

[0082] In step S1, a social network graph G of members is constructed s =(H,ε), where Let A∈R represent the original representation set of OG members, F represent the dimension, and ε represent the set of social relationship edges between OG members. m*m Representation graph G s The adjacency matrix of i and u j If there is a social relationship between ij The value in A is 1, otherwise, a ij The value of is 0.

[0083] Based on the degree of nodes, the shortest path length between nodes, and the length of edges in the social network graph, the social influence of each member in the temporary group is obtained. The formula is expressed as:

[0084]

[0085] Where INF(i) is social influence, R is the social network radius of the member, and R is set to 1; N x (i) is node u i x th The order of neighbors, d ij For node u i and node u j The shortest path length between j For node u j The degree of node u i The nth neighbor of ij is the weight of the social relationship between members, and ω is set ij =1.

[0086] The representation of members in OG is related to their different neighbors in the social network. In order to better learn the representation of members and incorporate the spatiotemporal preference information of members into the representation, we use the spatiotemporal attention graph neural network (abbreviated as ST-GAT) to learn from the graph G of OG. s ST-GAT is a variant of GNN, which combines the attention mechanism with the information propagation mechanism. It calculates the hidden state of each node based on the node's neighbors through the self-attention mechanism. In addition, ST-GAT uses multi-head attention to make the learning process more stable. Therefore, ST-GAT is suitable for learning the spatiotemporal information representation of members. The calculation process of member representation learning is described as follows:

[0087] First, for G s The initial representation of the node in is randomly initialized. Then, the member set representation H and the adjacency matrix A are imported into ST-GAT. ST-GAT updates the representation of the target node by collecting information about neighboring nodes. Let u i Representation graph G s Based on the spatiotemporal weight information between different nodes, we can obtain the spatiotemporal attention weights of different nodes:

[0088]

[0089] in, For node u i and u j The spatiotemporal attention weight, α is a hyperparameter used to balance the spatial and temporal relationship, which is a value between 0 and 1; LeakyReLU(·) is the activation function, and are the time weight matrix and the space weight matrix respectively, and The node u i and u j The time feature vector of and The node u i and u j Spatial eigenvector, i and j are the serial numbers of the nodes.

[0090] Furthermore, based on the spatiotemporal attention weights, the learnable weight matrix, the vector displacement and T, the comprehensive attention coefficient α is obtained ij , specifically expressed as:

[0091]

[0092] in, and For node u i and node u j Representation, W∈R F'*F is a learnable weight matrix, F is the dimension of the input feature, F' is the dimension of the output feature, || represents a cascade operation, for , R is the social network radius of the member, LeakyReLU(·) is the activation matrix, N i For node u i Neighbors, For node u i and u j The spatiotemporal attention coefficient of , exp(·) is the normalization coefficient of the Softmax function, and T is the transposed vector.

[0093] Furthermore, the normalized comprehensive attention coefficient α is obtained ij As node u i The new representation of is aggregated through ST-GAT according to the transformed representation of the neighboring nodes, which is specifically expressed as:

[0094]

[0095] in, is the representation of neighbor nodes after aggregation, σ is the nonlinear activation function, α ij is the comprehensive attention coefficient, is the neighbor node u j The representation of N i For node u i That is, member u i The neighbor set of , i and j are the serial numbers of the nodes.

[0096] In order to make the learning process of ST-GAT more stable, the present invention integrates a multi-head attention mechanism in ST-GAT. That is, the above transformation operation is performed by K independent attention mechanisms, and then the representations generated by each attention mechanism are connected to obtain the final representation of each node The process is described in detail as follows:

[0097]

[0098] Among them, σ is a nonlinear activation function, || represents a cascade operation, represents the normalized attention coefficient calculated based on the k-th comprehensive attention mechanism, W k represents the linearly changing weight matrix, Indicates member u i The final representation of is the neighbor node u j Representation of; By aggregating the information of neighboring nodes, the node u can be accurately captured. i That is, the social preferences of members are conducive to generating a fitting representation of OG.

[0099] For the social network graph G s For isolated nodes in , their representations cannot be updated during the forward propagation process due to their lack of neighbors. Therefore, the representations of isolated nodes can only be updated during the back-propagation process of ST-GAT.

[0100] In getting members After the member's social influence SI = [SI i ,SI j ,…,SI r] as weights, a weighted summation operation is used to generate a fitted representation of the temporary group OG:

[0101]

[0102] in, Represents the fitted feature representation of OG, which is generated based on the obtained member social influence and the member representation learned by ST-GAT, namely the spatiotemporal information representation.

[0103] In this step, in order to train the member representation learning module based on ST-GAT, we adopt the social network graph G s The edge prediction of is used as the driving task to activate the training of ST-GAT. Specifically, we first i and u j The representation between them is used to predict the probability of whether there is a social relationship between them, and then the loss of ST-GAT is calculated based on the predicted score and the ground truth value. Finally, ST-GAT is trained using the back propagation algorithm based on the loss. First, we calculate the predicted value y according to the MLP model pred .y pred represents the probability that there is a social relationship between two temporary group members, which is calculated as follows:

[0104]

[0105] Among them, MLP represents a neural network model consisting of two neural layers stacked together, || is used to connect vectors together, and Represents group members u i and u j y pred is a probability value in the range of (0, 1), indicating that member u i and u j Then, we choose binary cross entropy as the loss function, which is defined as the following equation:

[0106] τ(y pred )=y ture ·logy pred +(1-y ture )·log(1-y pred )

[0107] Among them, τ(y pred ) is the value of the loss function, y ture is the actual label value, which is 0 or 1; pred is a binary true value in the range of {0, 1}, and the loss function τ(y pred ) then, ST-GAT is based on τ(y pred) to perform back propagation mechanism.

[0108] S2: Learning the interaction preferences of members’ interest points: Based on the interaction records between members and interest points, we construct the connection edges between members, between members and interest points, and between interest points and interest points. Based on the popularity of interest points, we get the popularity weights of the connection edges between interest points and interest points, and build a bipartite graph of the interaction between members and interest points.

[0109] Based on popularity weights and learnable weight matrices, we can obtain information about the spread of members and POIs, and between POIs.

[0110] Based on the popularity weight, learnable parameter vector, transposed vector, learnable weight matrix, and spatiotemporal attention weight, the comprehensive popularity weight coefficient is obtained. The neighbor messages are further aggregated based on the comprehensive popularity weight coefficient to obtain the aggregated member messages and point of interest messages.

[0111] Multi-layer mapping is performed on the aggregated member messages and POI messages to update the members' interactive preference representation for POIs.

[0112] Specifically, OG = {u i ,u j ,…,u r} represents a temporary group, P G ={p1,p2,…,p k}, P G ∈P represents the set of POIs that OG members have visited, and I= i ,p j >(u i ∈OG,p j ∈P G ) represents the POI interaction record of the member. The bipartite graph of <member-POI> interaction incorporating the popularity of POI is constructed as follows: if u i Visited p j , then u i and p j construct a positive edge between them, otherwise, construct a negative edge between them; if u i and u m In p j If there are connections with the same sign on i and u m A positive edge is constructed between the two members, otherwise, a negative edge is constructed between the two members. In addition, according to the popularity of POI, a corresponding weight is assigned to each connecting edge to reflect the popularity of POI among different members.

[0113] Use G I =(OG,P G ,E​G ) represents an interactive bipartite graph, where OG and P G Represent the set of members and POIs respectively, E G Represents G I The connected edge set in E G =ε + +ε - , ε + and ε - Represent the positive edge and negative edge respectively. I As the input of the POI popularity enhanced bipartite graph neural network model (PEBGNN for short), the message propagation, message aggregation and update mechanism of PEBGNN is used to learn the representation of POI by aggregating the information of positive and negative neighbors. The specific process distribution is expressed as follows:

[0114] Message propagation: For the / th layer of SBGCN, and It is based on the positive and negative edges to transfer messages from p j Spread to u i , w pj Indicates POIp j The message propagation process between members and points of interest is expressed as:

[0115]

[0116] in, and The message is transferred from the interest point p based on the positive edge and the negative edge of the lth layer respectively. j Propagate to member u i , w pj Point of interest p j is the popularity weight, MSG is the message propagation function, is the member u at level l i The expression, is the interest point p at the lth layer j The expression, For the lth layer based on the positive edge from the interest point p j To member u i The spread of news, For the lth layer based on negative edges from the interest point p j To member u i of news spread.

[0117]

[0118] in, and The lth layer is based on positive edges and negative edges to send messages from member u toj Propagate to the point of interest p i , w pi Point of interest p i is the popularity weight, MSG is the message propagation function, is the member u at level l j The expression, is the interest point p at the lth layer i The expression, For the lth layer based on the positive edge from member u j To the point of interest i The spread of news, For the lth layer based on the negative edge from member u j To the point of interest i of news spread.

[0119] The message propagation process between points of interest is expressed as:

[0120]

[0121] in, and The message is transferred from the interest point p based on the positive edge and the negative edge of the lth layer respectively. j Propagate to the point of interest p i , w pj Point of interest p j is the popularity weight, MSG is the message propagation function, is the interest point p at the lth layer i The expression, is the interest point p at the lth layer j The expression, For the lth layer based on the positive edge from the interest point p j To the point of interest i The spread of news, For the lth layer based on negative edges from the interest point p j To the point of interest i of news spread.

[0122] Message aggregation: After obtaining the information of neighbor members, information aggregation is performed. The specific expression is as follows:

[0123]

[0124] Among them, α' ij is the comprehensive popularity weighted coefficient, W∈R F'*F is a learnable weight matrix, F is the dimension of the input feature, F' is the dimension of the output feature, || is a cascade operation, is the learnable parameter vector, LeakyReLU(·) is the activation matrix, N i For member ui The neighbor set of For member u i and member u j The spatiotemporal attention coefficient, w pj and w pk They are points of interest p j and points of interest k The popularity weight of , exp(·) is the Softmax function used to normalize the weight coefficient;

[0125]

[0126] Among them, N i is the neighbor set of node i, α' ij is the comprehensive popularity weighted coefficient, for node u i , whose message m l (u i ) is through u i For neighbor node u j Message l (u j ) is obtained by weighted summation; for node p i , whose message m l (p i ) is through p i For neighbor node p j Message l (p j ) is obtained by weighted summation.

[0127] Update mechanism: After the message is propagated and aggregated, each member u i Get four sets of information from its neighbor nodes, namely and Each point of interest p i It also obtains four sets of information from its neighboring nodes, namely and PEBGNN aggregates the information of four groups of neighboring nodes onto its own node and obtains the final representation of its own node through multi-layer mapping. The specific process is expressed as follows:

[0128]

[0129] in, is the representation of the member set, is the representation of the interest point set, MLP is a neural network model consisting of two fully connected neural layers stacked with loss rate and activation mechanism, w pj Point of interest p j The popularity weight, w pi Point of interest p i The popularity weight of is the representation of the updated l+1th layer member set, is the representation of the updated l+1th layer interest point set, and For each member u i Four sets of information obtained from neighbors, and For each point of interest p i Four sets of information obtained from neighbors.

[0130] MLP represents a neural network model consisting of two fully connected neural layers stacked with dropout and activation mechanisms:

[0131] MLP=W2(σ(dropout(W1x+b1)))+b2

[0132] Where W1, W2 and b1, b2 represent the learnable parameters of MLP. Specifically, W1 is the weight matrix of the first layer, W2 is the weight matrix of the second layer, b1 is the bias vector of the first layer, b2 is the bias vector of the second layer, σ is the activation function, and the dropout function is used to prevent overfitting.

[0133] The output sequence of PEBGNN is expressed as in is the representation of POI, that is, the member’s interest point p i The positive and negative POI interaction preferences are represented.

[0134] Furthermore, in S2, the interactive bipartite graph G I The edge prediction of is used as the driving task to activate the training of SBGNN. First, according to the member u i and points of interest i The representation between them is used to predict the probability of whether there is an interaction edge between them, and then the loss of SBGNN is calculated according to the predicted score and the ground truth value. Finally, the SBGNN is trained using the back propagation algorithm based on the loss.

[0135] First, calculate u according to the maximum likelihood estimation method i →p j The predicted values ​​are as follows:

[0136]

[0137] Where MLP represents a neural network model consisting of two stacked neural layers, || is used to connect vectors together, Indicates the representation of POI, Represents the representation of members, y pred is a probability value in the range of (0, 1), indicating that member u iand points of interest j The possibility of social relations between them.

[0138] We choose binary cross entropy as the loss function, which is expressed as follows:

[0139] τ=-ω[w pi ·y ture ·logy pred +(1-y ture )·log(1-y pred )]

[0140] Where τ is the loss function, ω is a weight parameter, and w pi For p i The popularity weight is used to overcome the imbalance problem of the number of positive and negative samples and make the training process of PEBGNN more stable; ture is the ground truth value mapped to {-1, 1} to {0, 1}.

[0141] S3: Learning members’ interest point transfer preferences: Based on the members’ visit order to interest points, a directed graph of interest point transfer is constructed with interest points as nodes and visit relationships between interest points as edges.

[0142] Based on the normalized weights of the connecting edges, the features of the interest points themselves, the weight matrix of the features of the interest points themselves, the features with contextual information transmitted from the neighboring interest points, the weight matrix of the features with contextual information transmitted from the neighboring interest points, and the bias vector, after weighted summation processing, the members' transfer preference representation for the interest points is obtained.

[0143] In the process of learning members' POI transfer preferences, we first construct a POI transfer directed graph based on the order of each member's visit to the POIs, that is, a directed graph is constructed based on each POI check-in sequence. Then, based on the POI transfer directed graph, a session-based graph convolutional neural network (SRGCN) is applied to learn the representation of POIs containing members' POI transfer preferences. In this process, the SRGCN model uses the node's own features and the contextual information with its neighboring nodes to update the node's embedded representation. Specifically, SRGCN takes the weighted sum of each node's features with those of its neighboring nodes and fuses the node's own features and contextual information through nonlinear transformation, so as to take into account the relationship between nodes in the learning process and capture members' POI transfer preferences.

[0144] Specifically, S = {S i ,S j ,…,S r} represents the check-in sequence set of all members in OG, where S i ={s1,s2,…,s i} represents member ui The sign-in sequence, and s i ={p1,p2,...p k}(k∈[0,n]) represents member u i The i-th POI sequence in the POI check-in sequence is based on the registration sequence s i ; Further construct the POI transfer directed graph Where P i represents the set of POIs, E i Representation diagram The set of connected edges in , edge e=(p k ,p j )(e∈E i ) represents member u i In POI sequence i In, visited p j After visiting p k .

[0145] A i ∈R n*2k Representation diagram The adjacency matrix of A. i Defined as two adjacency matrices and connection, it shows Weighted connections between edges, including incoming and outgoing edges. For each node, there is an incoming edge and outgoing edges Considering that multiple POIs may be displayed multiple times in a sequence, it is necessary to assign normalized weights to the connecting edges. The corresponding calculation formula is:

[0146]

[0147] in, is the normalized weight of the incoming edge connection, is the normalized weight of the outgoing edge connection, count(v i ,v j )∈{0,1}, count(v j ,v i ) both represent node v i and node v j Is there an edge between them? If yes, the value is 1; if no, the value is 0; count(v i ,v k )∈{0,1}, count(v k ,v i ) both represent node v i and node v kWhether there is an edge between them, if yes, the value is 1, if not, the value is 0; For node v i Predecessor node set, For node v i Post node set.

[0148] Furthermore, the POI transfer preferences of members are learned from the constructed POI transfer directed graph. represents the representation set of POI, where F represents the dimension; the representation set of POI Q i and the adjacency matrix A i Input into SRGCN to learn the transfer preference representation of members for interest points The formula is:

[0149]

[0150] in, Point of interest p i Combined with its neighboring interest points p j Characteristics at time t and its own characteristics x i , A ij From the point of interest p i To the point of interest j is the normalized weight of the connecting edge, P is the weight matrix with context information features passed from the neighboring interest points, Q is the weight matrix of the interest point’s own features, and b represents the deviation vector.

[0151] Further calculation of the update gate and reset gate values, the update gate controls what information should be retained, and the reset gate controls what information should be ignored. The process is as follows:

[0152]

[0153] in, is the update gate at time t, σ is the activation function, which compresses the input between 0 and 1 and is used for the gating signal; W z is the weight matrix associated with the current input, used to update the gate; is the current input at time step t, U z is the weight matrix associated with the state of the previous time step, used to update the gate; is the state at the previous time step t-1; is the reset gate at time step t, W r is the weight matrix associated with the current input, used to reset the gate, U r is the weight matrix associated with the state of the previous time step, used to reset the gate; represents the reset gate and σ(·) represents the activation function.

[0154] Node p i Candidate status According to its previous status Current Status and reset gate status Decision, specifically:

[0155]

[0156] Among them, W0 is the current input The associated weight matrix, U0 is the hidden state at the previous time step The associated weight matrix, is the current input at time step t.

[0157] Node p i The final state Based on the previous hidden state and candidate status Decision, in the update gate The calculation process in the calculation state is:

[0158]

[0159] After the learning process, the output sequence of SBCGN is expressed as in Indicates the p in POI i , that is, the members’ transfer preference representation for points of interest.

[0160] In this embodiment, in order to train the SRGCN model, sequence prediction is used as a driving task to activate the training of the SRGCN model. Specifically, the probability of each candidate POI is predicted, which indicates the possibility of a POI as the last interactive POI in the sequence. If the predicted POI probability is large, it means that the POI is more likely to be the last interactive POI in the sequence. Then, the loss of SRGCN is calculated based on the prediction score and the ground truth value, and finally the SRGCN is trained using the back propagation algorithm based on the calculated loss.

[0161] Consider the POI transmission sequence s i Local embedding of For POI transfer sequence s i ={p1,p2,...p k}, local embedding Can be defined as the last check-in POIp n of We further consider the POI transmission sequence s by aggregating the representations of all POIs i Global embedding of Considering that the information involved in these POI representations may have different degrees of priority, a soft attention mechanism is used to better describe the global embedding This is shown in the following equation:

[0162]

[0163] Among them, α i is the attention coefficient for node i, is the scaling factor, is a constant, T is the exponent of the scaling factor, For about The weight matrix is ​​used to transform the feature vector of node n Mapping to a common feature space; For about The weight matrix is ​​used to transform the feature vector of node i Mapped to a common feature space, c is a trainable parameter.

[0164] Second, the hybrid embedding is computed by linearly transforming the concatenation of the local and global embedding representations The matrix W3 is responsible for and Compressed to the embedding space, the formula is as follows:

[0165]

[0166] Among them, W3 is the weight matrix, which is used to map the concatenated vector to the new feature space; is the local embedding of the last checked POI in the POI transfer sequence, is a global embedding generated by aggregating the representations of all POIs in the sequence.

[0167] Furthermore, after obtaining each POI transfer sequence After the embedding representation, we represent the feature vector of node i, i.e. the representation vector of the candidate POI, by Multiply by mixed embedding To calculate each candidate POIp i Score The specific calculation formula is:

[0168]

[0169] Finally, among all the prediction scores Construct a prediction vector on the cascade of And use the normalization function sotfmax to obtain an output vector The formula is as follows:

[0170]

[0171] Where σ is the activation function, is the prediction score of all candidate POI sets, is the score of the nth candidate POI, Transmit sequence s for POI display to POI i The probability of the last interaction.

[0172] For each POI transfer sequence, based on the prediction and the basic facts, a loss function is defined according to the cross entropy function. It is expressed as the formula:

[0173]

[0174] in One-hot encoded vector representing the true POI, is the predicted probability of the i-th POI, is the actual label of the i-th POI. It should be noted that the length of most POI transfer sequences is relatively short. Therefore, it is recommended to choose relatively few training cycles to prevent overfitting.

[0175] S4: Recommend the next point of interest: The members' interaction preference representations for the points of interest and the members' transfer preference representations for the points of interest are summed to obtain a final point of interest representation, and a vector product operation and a normalization operation are performed on the final point of interest representation and the fitting feature representation of the temporary group to obtain the score of the temporary group on each candidate point of interest, and the order of recommending the points of interest to the temporary group is determined based on the score.

[0176] After performing vector product operations and normalization operations on the final interest point representation and the fitting feature representation of the temporary group, the score of the temporary group on each candidate interest point is obtained, which is expressed as follows:

[0177]

[0178] in, is the final POI representation, Z is the POI interaction preference representation, S is the POI transfer preference representation, is the member’s POI interaction preference representation, A member's POI transfer preference representation.

[0179]

[0180] in, For each interest point, OG score is given, and Sigmoid is the normalization operation function. is the final point of interest representation, is the fitting characteristic representation of the temporary group, is the transpose of the fitted feature representation for the temporary group.

[0181] The present invention also provides a temporary group next point of interest recommendation system based on hybrid graph neural network, comprising:

[0182] Temporary group fitting feature representation module: used to obtain the social relationships of members in the temporary group, build a social network graph, and obtain the social influence of each member in the temporary group based on the degree of nodes, the shortest path length between nodes, and the length of edges in the social network graph; used to obtain the spatiotemporal attention weight of the node based on the temporal feature vector, spatial feature vector, temporal weight matrix, and spatial weight matrix of the node; based on the spatiotemporal attention weight, learnable weight matrix, vector displacement, and transposed vector, obtain the comprehensive attention coefficient, and generate the spatiotemporal information representation of the node after performing aggregation and multi-head attention operations on neighbor nodes in turn based on the comprehensive attention coefficient; used to obtain the fitting feature representation of the temporary group after weighted summation based on the obtained social influence and spatiotemporal information representation;

[0183] Learning member interest point interaction preference module: based on the interaction records between members and interest points, construct connection edges between members and members, between members and interest points, and between interest points and interest points, obtain the popularity weights of the connection edges between interest points and interest points based on the popularity of interest points, construct an interactive bipartite graph of members and interest points, obtain the message propagation between members and interest points and interest points and interest points based on the popularity weights and the learnable weight matrix, obtain the comprehensive popularity weighting coefficient based on the popularity weights, learnable parameter vectors, transposed vectors, learnable weight matrices, and spatiotemporal attention weights, further aggregate neighbor messages based on the comprehensive popularity weighting coefficient, obtain aggregated member messages and interest point messages, respectively, perform multi-layer mapping processing on the aggregated member messages and interest point messages, and update to obtain the member's interaction preference representation for interest points;

[0184] Learning member interest point transfer preference module: used to construct an interest point transfer directed graph with interest points as nodes and access relations between interest points as edges based on the member's visit order to interest points.

[0185] Based on the normalized weights of the connecting edges, the features of the interest points themselves, the weight matrix of the features of the interest points themselves, the features with context information transmitted from the neighboring interest points, the weight matrix of the features with context information transmitted from the neighboring interest points, and the bias vector, after weighted summation processing, the transfer preference representation of the members for the interest points is obtained;

[0186] Recommendation module: used to obtain the final interest point representation by summing up the members' interaction preference representations for interest points and the members' transfer preference representations for interest points, and to obtain the score of the temporary group on each candidate interest point after performing vector product operation and normalization operation on the final interest point representation and the fitting feature representation of the temporary group, and to determine the order of recommending interest points to the temporary group based on the score.

[0187] Experimental Section

[0188] 1. Dataset and Experimental Environment

[0189] In the research of next POI recommendation, the existing public datasets are mainly for individual users or fixed groups, and there is a lack of dedicated datasets for next POI recommendation for temporary groups. In order to solve this problem, a method for next point of interest recommendation for temporary groups based on hybrid graph neural network (abbreviated as STC-NPROG) is proposed. First, the temporary group is fitted into a virtual user, and then the next POI recommendation for the temporary group is converted into personalized next POI recommendation for the virtual user. In the experimental part, three sets of personalized next POI recommendation datasets are selected to verify the performance of the STC-NPROG model, namely Foursquare, Gowalla and Yelp. These three datasets contain rich user social data and POI interaction data. Each user POI interaction data records the user's ID and POI information in detail. In order to ensure the accuracy and effectiveness of the experiment and keep consistent with the experimental settings of the baseline model, the baseline model is referred to in the experiment, and data preprocessing operations are performed on the original dataset to filter out POIs with a visit frequency of less than 10 times and users with less than 10 check-ins. The information of the filtered dataset is shown in Table 1. After performing data preprocessing, in each dataset, 60% of the data is randomly selected as a training set, 20% of the data is used as a validation set, and the remaining 20% ​​of the data is used as a test set.

[0190] Temporary groups are constructed by randomly selecting a group of users with social relationships from the dataset, and the social relationships between some members of the group and other members are deleted to simulate isolated members without social relationships. In addition, the POI features containing the POI interaction preferences of group members are learned based on the POI interaction data of group members, and the POI features containing the POI transfer preferences of group members are learned based on the POI transfer data of group members.

[0191] Table 1 Statistics of Foursquare, Gowalla and Yelp datasets

[0192]

[0193] The experimental operating environment of the temporary group point of interest recommendation model based on hybrid graph neural network (abbreviated as STC-NPROG) is as follows: the GPU uses the NVIDIA GeForce RTX 3090 graphics card to ensure the efficiency of model training and reasoning; the operating system uses Windows 10 Professional 64-bit to provide a stable computing environment; the memory configuration is 32GB RAM to meet the memory requirements of large neural network models; the CPU is the 12th generation Intel Core i9-12900K, with a main frequency of 3.19GHz, providing powerful computing power; the deep learning model framework uses Pytorch based on Python 3.8 version to support model building, training and deployment; the development environment uses the Pycharm integrated development environment to provide a convenient programming and debugging experience

[0194] 2. Experimental evaluation

[0195] 2.1 Recommendation Model Performance Evaluation Indicators

[0196] In order to evaluate the performance of the STC-NPROG model, we used three widely used evaluation indicators in the experiment, namely, precision (P@K), recall (R@K) and average precision (MAP) (M@K). The formulas of the three indicators are defined as follows:

[0197]

[0198] Among them, R(g) represents the POI recommendation list, T(g) represents the real interaction list of the temporary group OG. In this paper, N represents the length of the POI recommendation list, k' is the number of the first few results to be evaluated, which is a fixed value; K' is the total number of results to be evaluated, and the value of K' is set at K'=[2,5,10,20].

[0199] 2.2 Baseline Model

[0200] Specifically, 11 personalized next POI recommendation models were selected as baseline models to conduct performance comparison experiments with STC-NPROG to evaluate and verify the recommendation performance of STC-NPROG. The 11 baseline models are introduced as follows:

[0201] KYD: KYD is a hybrid deep learning model based on graph convolutional networks and attention mechanism.

[0202] TDGCN: TDGCN is a POI recommendation model based on the knowledge graph attention network, which uses the graph attention mechanism to capture the complex relationship between users and POIs.

[0203] ATST-GGNN: ATST-GGNN is a spatiotemporal gating graph neural network model based on attention mechanism.

[0204] GETNext: GETNext uses user check-in sequences and global trajectory flow graphs, combines user preferences and spatiotemporal context information, and constructs a graph neural network and attention mechanism for prediction.

[0205] SUCP: SUCP is a point of interest (POI) recommendation model based on user activity centers and social influence. It combines geographic, social and time information through matrix decomposition technology to predict users' POI preferences.

[0206] LSTPM: LSTPM adopts non-local network and geographically expanded RNN techniques to model users' long-term and short-term preferences, respectively.

[0207] STP-UDGAT: STP-UDGAT,predicts users’ POI preferences by combining geographical, temporal and user preference information through,graph attention network.

[0208] SGRec: SGRec is a point-of-interest (POI) recommendation model based on user behavior sequences and graph structures, which combines geographic, temporal and user preference information for prediction through a graph attention network (GAT).

[0209] TMCA: TMCA is a POI recommendation model based on user behavior sequence and multi-level contextual attention mechanism. It adopts the LSTM encoder-decoder framework to automatically learn the spatiotemporal representation of historical check-in activities.

[0210] NHRM: NHRM is a point of interest (POI) recommendation model based on user characteristics and spatiotemporal factors. The model consists of three sub-models, which consider user preferences, forgotten features, user influence, trajectory, POI location correlation and POI category influence respectively.

[0211] AGRAN: AGRAN is a point-of-interest (POI) recommendation model based on adaptive graph representation learning. It enhances the attention network through adaptive graph representation and combines geographic, temporal and user preference information to predict users' POI preferences.

[0212] 2.3 Impact of Group Size on Recommendation Performance

[0213] For group recommendation, group size is one of the important factors affecting the recommendation performance of the model. This section conducts an experimental study on the impact of group size on the recommendation performance of STC-NPROG. Specifically, we first set the group size interval to [10, 20, 30, 40, 50, 60, 70, 80, 90, 100]; then, we use Precision@K and NDCG@K as performance evaluation indicators to test the recommendation performance of STC-NPROG under different group sizes. The experimental results are shown in Figure 2. Figure 2 shown.

[0214] from Figure 2 We can find that the mean precision (Mean Precision) of the STC-NPROG model fluctuates with the increase of group size on the three datasets. In addition, the performance of STC-NPROG on the Gowalla and Yelp datasets is weaker than that on the Foursquare dataset. The experimental results are shown in the figure. Specifically, on the Foursquare dataset, when the group size increases from 20 to 70, the Mean Precision value first increases from 0.86 to 0.91, and then decreases to 0.84. Afterwards, when the group size increases from 70 to 100, the Mean Precision value gradually increases to 0.91. On the Gowalla dataset, the Mean Precision value first stabilizes at around 0.80, then increases to 0.87 when the group size increases from 60 to 90, and then stabilizes at 0.86. In the Yelp dataset, when the group size increases from 10 to 40, the Mean Precision value first decreases from 0.08 to 0.05, and then increases to 0.1 as the group size increases from 50 to 60. After that, the value of Mean Precision first decreases and then increases to 0.1 as the group size changes.

[0215] Therefore, when the group size is small, the lack of training data will limit the improvement of STC-NPROG's recommendation performance. However, the continued growth of the group size will also affect the recommendation performance of STC-NPROG. Therefore, in the related research on the next POI recommendation of the group, the group size does have an impact on the recommendation performance of the model, and choosing the appropriate group size is crucial to obtaining the optimal recommendation performance of the model. In this chapter, the optimal group size of STC-NPROG is set to 50, 90, and 100 on the three datasets of Foursquare, Gowalla, and Yelp, respectively. This setting helps STC-NPROG achieve the best performance on these three datasets.

[0216] 2.4 Performance Comparison

[0217] In order to verify the recommended performance of STC-NPROG, this experiment uses Precision@K and NDCG@K as performance evaluation indicators, and conducts performance comparison experiments between STC-NPROG and the baseline model. It should be pointed out that since some key parameters are not clearly given in the original papers corresponding to the baseline models, it is difficult for this invention to accurately reproduce the experimental results of the baseline model. In order to ensure the fairness of the comparative experiment as much as possible, this chapter uses the experimental results of each baseline model in the original paper as the experimental results of the baseline model, and the experimental results of STC-NPROG are derived from the calculation and operation of the model. The final experimental results are shown in Table 2, Table 3 and Table 4. In the above three tables, the best experimental data under different indicators are bolded. In particular, since the values ​​of @E of each baseline model are not exactly the same as those of STC-NPROG, there are missing data in Table 2, Table 3 and Table 4. This chapter uses the "—" symbol to specially mark the missing data.

[0218] On the Foursquare dataset, AGRAN, LSTPM, ATST-GGNN, STP-UDGAT, SGRec, TDGCN, GETNext and KYD were selected as baseline models, and performance comparison experiments were carried out with the STC-NPROG model. The experimental results are shown in Table 2. As can be seen from Table 2, for Precision@K and NDCG@K, our proposed model STC-NPROG outperforms other baseline models overall. Specifically, when K increases from 2 to 20, Precision@K first increases from 0.96 to 0.9, and then gradually decreases to 0.84, while NDCG@K decreases from 0.95 to 0.86 as K increases.

[0219] KYD performs better than the baseline models in the Foursquare dataset. Therefore, compared with the KYD model, the STC-NPROG model improves Precision@K by 24.2%, and NDCG@K by 57.2% compared with all baseline models.

[0220] Table 2 Performance comparison of STC-NPROG model and baseline model on Foursquare dataset

[0221]

[0222] On the Gowalla dataset, the present invention selected SUCP, SGRec, LSTPM, AGRA, ATST-GGNN, GETNext, STP-UDGAT, TDGCN and KYD as baseline models, and conducted performance comparison experiments with the STC-NPROG model. The experimental results are shown in Table 3. As can be seen from Table 3, for Precision@K and NDCG@K, the STC-NPROG we use is generally better than other baseline models. Specifically, when K increases from 2 to 20, Precision@K remains basically stable, fluctuating around 0.85. And NDCG@K also fluctuates slightly, basically remaining around 0.86.

[0223] KYD performs better than the baseline models in the Gowalla dataset, but the KYD model only has data for Precision@K. Therefore, compared with the KYD model, STC-NPROG improves Precision@K by 35%, and NDCG@K improves by 63% compared with all baseline models.

[0224] Table 3 Performance comparison results of STC-NPROG model and baseline model on Gowalla dataset

[0225]

[0226] In the research field of POI recommendation, the frequency of use of the Yelp dataset is lower than that of the Foursquare and Gowalla datasets. The baseline models selected in this paper are TMCA, SUCP and NHRM, and the performance comparison experiments are carried out with the STC-NPROG model. The experimental results are shown in Table 4. As can be seen from Table 4, for Precision@K and NDCG@K, the STC-NPROG we use is better than other baseline models on some data. Specifically, when K increases from 2 to 5, Precision@K decreases from 0.1 to 0.05, and when K increases from 5 to 20, Precision@K increases from 0.05 to 0.1. NDCG@K gradually decreases with the increase of K value and then tends to be stable. When K increases from 2 to 20, NDCG@KK decreases from 0.1 to 0.07 and then increases to 0.09. However, NDCG@K is lower than the NHRM model when K increases from 10 to 20.

[0227] NHRM performs better than the baseline models in the Yelp dataset. Therefore, compared with the NHRM model, the STC-NPROG model improves Precision@K by 12.5%.

[0228] Table 4 Comparison results of STC-NPROG model and baseline model on Yelp dataset

[0229]

[0230] 2.5 Ablation Experiment

[0231] STC-NPROG includes four main modules, namely, group feature learning module, user POI interaction preference learning module, user POI transfer preference learning module and POI recommendation module. Among the above four modules, the user POI interaction preference learning module and the POI transfer preference learning module can be removed in the STC-NPROG structure, so as to explore the performance of the model after simplifying the structure, and further analyze the influence of these two modules on the overall performance of STC-NPROG. By removing these two modules, the role of each component in STC-NPROG can be more clearly understood, providing new ideas for model optimization and performance improvement. However, since the group feature learning module is responsible for generating the fitting features of the temporary group, and the POI recommendation module is responsible for generating the next POI recommendation result of the temporary group, the lack of the above modules will cause the calculation process of STC-NPROG to fail to execute normally. Therefore, the present invention no longer carries out ablation experiments for this module. Based on the structure of STC-NPROG, the present invention constructs the following two groups of ablation models:

[0232] POI popularity enhanced bipartite graph neural network ablation model (PEBGNN-NPROG): Based on the STC-NPROG structure, this model removes the group member POI transfer preference learning module and only uses the group feature fitting module and the group member POI interaction preference learning module to generate the next POI recommendation result for the temporary group.

[0233] Session-based graph convolutional neural network ablation model (SRGCN-NPROG): Based on the STC-NPROG structure, this model removes the group member POI interaction preference learning module and only uses the group feature fitting module and the group member POI transfer preference learning module to generate the next POI recommendation result for the temporary group.

[0234] In this section, Precision@K and NDCG@K are used as performance evaluation indicators, and ablation experiments are carried out on STC-NPROG, PEBGNN-NPROG and SRGCN-NPROG on the Foursquare dataset to explore the impact of different learning modules on the model recommendation performance in the model structure of STC-NPROG. The experimental results are shown in Table 5.

[0235] Table 5 Ablation experiment results on Foursquare dataset

[0236]

[0237] From the experimental results in Table 5, we can see that on the Foursquare dataset, SRGCN-NPROG has the lowest Precision@K and NDCG@K values, while PEBGNN-NPROG has better Precision@K and NDCG@K values ​​than SRGCN-NPROG, but weaker than STC-NPROG. The influence of user's POI transfer preference on the performance of STC-NPROG is weaker than that of user's POI interaction preference. This finding provides an important reference for further improving and optimizing STC-NPROG. In addition, the experimental results in Table 5 show that the Precision@K and NDCG@K values ​​of STC-NPROG on the Foursquare dataset are better than those of the other two groups of ablation models, indicating that it is reasonable for STC-NPROG to consider user's POI interaction preference and POI transfer preference when recommending the next POI for temporary groups. This design helps STC-NPROG to learn user behavior patterns and interaction preferences more comprehensively, thereby further improving the quality of the next POI recommendation for temporary groups.

[0238] 2.6 Experimental verification of the feature fitting model integrating spatiotemporal information

[0239] This paper takes the spatiotemporal information preferences of group members into consideration for the first time, and proposes a new feature fitting method (ST-OGRF) for temporary groups that integrates spatiotemporal information features. The core point of ST-OGRF is to generate fitting features for temporary groups by integrating the social influence, social network relationships and spatiotemporal dynamic information of group members through the spatiotemporal graph attention network (ST-GAT) and node influence index (INF). In order to verify the effectiveness of the ST-OGRF model, a comparative model of STC-NPROG was first constructed as a feature fitting model (NPROG-OGRF) that integrates spatiotemporal information. The construction of NPROG-OGRF is based on the structure of STC-NPROG, removing the operation of calculating spatiotemporal information of group members, and learning the fitting features of temporary groups only based on the social influence and social network relationships of members; then, Precision@K and NDCG@K were used as performance evaluation indicators, and the value of K was set at K = [2, 5, 10, 15, 20]. Performance comparison experiments were carried out on STC-NPROG and NPROG-OGRF on the Foursquare dataset. The experimental results are shown in Table 6.

[0240] Table 6 Comparison of the impact of feature fitting integrating spatiotemporal information on model recommendation performance

[0241]

[0242] As can be seen from Table 6, on the Gowalla dataset, the recommendation performance of STC-NPROG is better than that of NPROG-OGRF. In terms of Precision@K and NDCG@K, STC-NPROG has improved by 35.2% and 30.2% respectively. The experimental comparison results show that the operation of incorporating the spatiotemporal information preferences of group members is reasonable and effective, which is conducive to STC-NPROG to learn the fitting features of temporary groups more accurately, thereby further improving the recommendation effect of the model.

[0243] 2.7 Experimental Verification of POI Popularity Learning Mechanism

[0244] In order to learn the POI interaction preferences of group members more comprehensively, this paper proposes a POI popularity enhanced bipartite graph neural network model (PEBGNN). Different from the existing bipartite graph neural network, PEBGNN not only learns the node information in the interactive bipartite graph, but also learns the POI popularity, which makes PEBGNN have a more powerful feature learning ability. In order to verify the effectiveness of the POI popularity learning mechanism, this section conducts experimental verification on the POI popularity learning mechanism. Specifically, firstly, a comparison model (NPROG-SBGNN) without the POI popularity learning mechanism is constructed based on STC-NPROG, and the remaining modules are consistent with STC-NPROG; then, Precision@K and NDCG@K are used as performance evaluation indicators, and the value of K is set at K=[2,5,10,15,20]. Performance comparison experiments are carried out on STC-NPROG and NPROG-SBGNN on the Gowalla dataset. The experimental results are shown in Table 7.

[0245] Table 7 Comparison of the impact of POI popularity learning mechanism on model recommendation performance

[0246]

[0247] As can be seen from Table 7, on the Gowalla dataset, the recommendation performance of STC-NPROG is better than that of NPROG-SBGNN. STC-NPROG improves Precision@K and NDCG@K by 3% and 6.2% respectively. The experimental results show that it is reasonable and effective to introduce the POI popularity learning mechanism in STC-NPROG, which helps STC-NPROG to learn the POI interaction preferences of group members more comprehensively, thereby further improving the quality of the next POI recommendation for temporary groups.

[0248] 2.8 Experimental Verification of Learning Transfer Preferences by Incorporating Contextual Information

[0249] In order to learn the POI transfer preferences of group members more comprehensively, this paper proposes a graph convolutional neural network model (SRGCN) based on contextual information learning. Unlike the existing graph neural network based on session, SRGCN not only learns the node information in the transition graph of POI, but also learns the contextual information of the node, which makes SRGCN have a more powerful feature learning ability. In order to verify the effectiveness of learning the transfer preference mechanism with contextual information, this section conducts experimental verification on the learning transfer preference mechanism with contextual information. Specifically, firstly, a comparison model (NPROG-SRGNN) is constructed based on STC-NPROG to remove the learning transfer preference mechanism with contextual information, and the remaining modules are consistent with STC-NPROG; then, Precision@K and NDCG@K are used as performance evaluation indicators, and the value of K is set at K=[2,5,10,15,20]. Performance comparison experiments are carried out on STC-NPROG and NPROG-SRGNN on the Gowalla dataset. The experimental results are shown in Table 8.

[0250] Table 8 Comparison of the impact of learning transfer preference based on contextual information on model recommendation performance

[0251]

[0252] As can be seen from Table 8, on the Gowalla dataset, the recommendation performance of STC-NPROG is better than that of NPROG-SRGNN. STC-NPROG improves Precision@K and NDCG@K by 8.2% and 8.4% respectively. The experimental results show that it is reasonable and effective to introduce the transfer preference mechanism combined with context information in STC-NPROG. This mechanism helps STC-NPROG to learn the POI transfer preferences of group members more comprehensively, thereby further improving the quality of the next POI recommendation for temporary groups.

Claims

1. A method for recommending the next point of interest in a temporary group based on a hybrid graph neural network, characterized in that: The following steps are involved: S1. Temporary group fitting feature representation: obtain the social relationships of members in the temporary group, build a social network graph, and obtain the social influence of each member in the temporary group based on the degree of nodes, the shortest path length between nodes, and the length of edges in the social network graph; Based on the node's temporal feature vector, spatial feature vector, temporal weight matrix and spatial weight matrix, the spatiotemporal attention weight of the node is obtained; Based on the spatiotemporal attention weight, learnable weight matrix, vector displacement and transposed vector, the comprehensive attention coefficient is obtained. After the neighbor nodes are aggregated and multi-head attention operations are performed in sequence based on the comprehensive attention coefficient, the spatiotemporal information representation of the node is generated; Based on the obtained social influence and spatiotemporal information representation, a weighted sum is performed to obtain the fitting feature representation of the temporary group; S2. Learning the interaction preferences of members’ interest points: Based on the interaction records between members and interest points, we construct the connection edges between members, between members and interest points, and between interest points and interest points. Based on the popularity of interest points, we get the popularity weights of the connection edges between interest points and interest points, and build a bipartite graph of the interaction between members and interest points. Based on popularity weights and learnable weight matrices, we can obtain information about the spread of members and POIs, and between POIs. Based on the popularity weight, learnable parameter vector, transposed vector, learnable weight matrix, and spatiotemporal attention weight, the comprehensive popularity weight coefficient is obtained. The neighbor messages are further aggregated based on the comprehensive popularity weight coefficient to obtain the aggregated member messages and point of interest messages. Perform multi-layer mapping processing on the aggregated member messages and POI messages, and update the member's interactive preference representation for POIs; S3. Learning members’ interest point transfer preferences: Based on the members’ visit order to interest points, a directed graph of interest point transfer is constructed, with interest points as nodes and visit relationships between interest points as edges. Based on the normalized weights of the connecting edges, the features of the interest points themselves, the weight matrix of the features of the interest points themselves, the features with context information transmitted from the neighboring interest points, the weight matrix of the features with context information transmitted from the neighboring interest points, and the bias vector, after weighted summation processing, the transfer preference representation of the members for the interest points is obtained; S4. Recommend the next point of interest: sum the members' interaction preference representations for the points of interest and the members' transfer preference representations for the points of interest to obtain a final point of interest representation, perform vector product operations and normalization operations on the final point of interest representation and the fitting feature representation of the temporary group, obtain the score of the temporary group on each candidate point of interest, and determine the order of recommending the points of interest to the temporary group based on the scores.

2. The method for recommending the next point of interest in a temporary group based on a hybrid graph neural network according to claim 1, characterized in that: In S3, the normalized weight of the connecting edge, the feature of the point of interest itself, the weight matrix of the feature of the point of interest itself, the feature with context information transmitted from the neighboring point of interest, the weight matrix of the feature with context information transmitted from the neighboring point of interest, and the bias vector are weighted and summed to obtain the member's transfer preference representation for the point of interest. Specifically: in, Point of interest p i Combined with its neighboring interest points p j Characteristics at time t and its own characteristics x i , A ij From the point of interest p i To the point of interest j is the normalized weight of the connecting edge, P is the weight matrix with context information features passed from the neighboring interest points, Q is the weight matrix of the interest point’s own features, and b represents the bias vector.

3. The method for recommending the next point of interest in a temporary group based on a hybrid graph neural network according to claim 1, characterized in that: In S2, based on the popularity weight and the learnable weight matrix, the message propagation between members and points of interest, and between points of interest is obtained, which is specifically expressed as: in, and The lth layer is based on positive edges and negative edges to transfer messages from the point of interest p to j Propagate to member u i , w pj Point of interest p j is the popularity weight, MSG is the message propagation function, is the member u at level l i The expression, is the interest point p at the lth layer j The expression, For the lth layer based on the positive edge from the interest point p j To member u i The spread of news, For the lth layer based on negative edges from the interest point p j To member u i the spread of news; in, and The lth layer is based on positive edges and negative edges to send messages from member u to j Propagate to the point of interest p i , w pi Point of interest p i is the popularity weight, MSG is the message propagation function, is the member u at level l j The expression, is the interest point p at the lth layer i The expression, For the lth layer based on the positive edge from member u j To the point of interest i The spread of news, For the lth layer based on the negative edge from member u j To the point of interest i the spread of news; in, and The lth layer is based on positive edges and negative edges to transfer messages from the point of interest p to j Propagate to the point of interest p i , w pj Point of interest p j is the popularity weight, MSG is the message propagation function, is the interest point p at the lth layer i The expression, is the interest point p at the lth layer j The expression, For the lth layer based on the positive edge from the interest point p j To the point of interest i The spread of news, For the lth layer based on negative edges from the interest point p j To the point of interest i of news spread.

4. The method for recommending the next point of interest in a temporary group based on a hybrid graph neural network according to claim 1, characterized in that: In S2, a comprehensive popularity weight coefficient is obtained based on the popularity weight, the learnable parameter vector, the transposed vector, the learnable weight matrix, and the spatiotemporal attention weight. Neighborhood message aggregation is further performed based on the comprehensive popularity weight coefficient to obtain aggregated member messages and point of interest messages, respectively, which are specifically expressed as follows: Among them, α' ij is the comprehensive popularity weighted coefficient, W∈R F'*F is a learnable weight matrix, F is the dimension of the input feature, F' is the dimension of the output feature, || is a cascade operation, is the learnable parameter vector, LeakyReLU(·) is the activation matrix, N i For member u i The neighbor set of For member u i and member u j The spatiotemporal attention coefficient, w pj and w pk They are points of interest p j and points of interest k The popularity weight of , exp(·) is the Softmax function used to normalize the weight coefficient; Among them, N i is the neighbor set of node i, α' ij is the comprehensive popularity weighted coefficient, for node u i , whose message m l (u i ) is through u i For neighbor node u j Message l (u j ) is obtained by weighted summation; for node p i , whose message m l (p i ) is through p i For neighbor node p j Message l (p j ) is obtained by weighted summation.

5. The method for recommending the next point of interest in a temporary group based on a hybrid graph neural network according to claim 1, characterized in that: In S2, the aggregated member messages and POI messages are subjected to multi-layer mapping processing, and the member's interaction preference representation for POI is updated, which is specifically expressed as: in, is the representation of the member set, is the representation of the interest point set, MLP is a neural network model consisting of two fully connected neural layers stacked with loss rate and activation mechanism, w pj Point of interest p j The popularity weight, w pi Point of interest p i The popularity weight of is the representation of the updated l+1th layer member set, is the representation of the updated l+1th layer interest point set, and For each member u i Four sets of information obtained from neighbors, and For each point of interest p i Four sets of information obtained from neighbors.

6. The method for recommending the next point of interest in a temporary group based on a hybrid graph neural network according to claim 1, characterized in that: In S1, the spatiotemporal attention weight of the node is obtained based on the temporal feature vector, spatial feature vector, temporal weight matrix and spatial weight matrix of the node. The formula is: in, For node u i and u j The spatiotemporal attention weight, α is a hyperparameter used to balance the spatial and temporal relationship, which is a value between 0 and 1; LeakyReLU(·) is the activation function, and are the time weight matrix and the space weight matrix respectively, and The node u i and u j The time feature vector of and The node u i and u j Spatial eigenvector, i and j are the serial numbers of the nodes.

7. The method for recommending the next point of interest in a temporary group based on a hybrid graph neural network according to claim 1, characterized in that: In S1, the comprehensive attention coefficient α is obtained based on the spatiotemporal attention weight, the learnable weight matrix, the vector displacement and the transposed vector. ij , specifically expressed as: in, and For node u i and node u j Representation, W∈R F’*F is a learnable weight matrix, F is the dimension of the input feature, F' is the dimension of the output feature, || represents a cascade operation, for , R is the social network radius of the member, LeakyReLU(·) is the activation matrix, N i For node u i Neighbors, For node u i and u j The spatiotemporal attention coefficient of , exp(·) is the normalization coefficient of the Softmax function, and T is the transposed vector.

8. The method for recommending the next point of interest in a temporary group based on a hybrid graph neural network according to claim 1, characterized in that: In S1, the social influence of each member in the temporary group is obtained based on the degree of the nodes in the social network graph, the shortest path length between nodes, and the length of the edge. The formula is expressed as: Where INF(i) is social influence, R is the social network radius of the member, and R is set to 1; N x (i) is node u i x th The order of neighbors, d ij For node u i and node u j The shortest path length between j For node u j The degree of x is the node u i The nth neighbor of ij is the weight of the social relationship between members, and ω is set ij =1.

9. The method for recommending the next point of interest in a temporary group based on a hybrid graph neural network according to claim 1, characterized in that: The normalized weight of the connecting edge in S3 is expressed as follows: in, is the normalized weight of the incoming edge connection, is the normalized weight of the outgoing edge connection, count(v i ,v j )∈{0,1}, count(v j ,v i ) both represent node v i and node v j Is there an edge between them? If yes, the value is 1; if no, the value is 0; count(v i ,v k )∈{0,1}, count(v k ,v i ) both represent node v i and node v k Whether there is an edge between them, if yes, the value is 1, if not, the value is 0; For node v i Predecessor node set, For node v i Post node set.

10. A temporary group next point of interest recommendation system based on hybrid graph neural network, characterized in that: include: Temporary group fitting feature representation module: used to obtain the social relationships of members in the temporary group, build a social network graph, and obtain the social influence of each member in the temporary group based on the degree of the nodes, the shortest path length between nodes, and the length of the edges in the social network graph; Used to obtain the spatiotemporal attention weight of the node based on the node's temporal feature vector, spatial feature vector, temporal weight matrix and spatial weight matrix; Based on the spatiotemporal attention weight, learnable weight matrix, vector displacement and transposed vector, a comprehensive attention coefficient is obtained. After the neighbor nodes are aggregated and multi-head attention operations are performed in sequence based on the comprehensive attention coefficient, the spatiotemporal information representation of the node is generated; it is used to obtain the fitting feature representation of the temporary group after weighted summation based on the obtained social influence and spatiotemporal information representation; Learning member interest point interaction preference module: based on the interaction records between members and interest points, construct connection edges between members and members, between members and interest points, and between interest points and interest points, obtain the popularity weights of the connection edges between interest points and interest points based on the popularity of interest points, construct an interactive bipartite graph of members and interest points, obtain the message propagation between members and interest points and interest points and interest points based on the popularity weights and the learnable weight matrix, obtain the comprehensive popularity weighting coefficient based on the popularity weights, learnable parameter vectors, transposed vectors, learnable weight matrices, and spatiotemporal attention weights, further aggregate neighbor messages based on the comprehensive popularity weighting coefficient, obtain aggregated member messages and interest point messages, respectively, perform multi-layer mapping processing on the aggregated member messages and interest point messages, and update to obtain the member's interaction preference representation for interest points; Learning member interest point transfer preference module: used to construct an interest point transfer directed graph with interest points as nodes and access relations between interest points as edges based on the member's visit order to interest points. Based on the normalized weights of the connecting edges, the features of the interest points themselves, the weight matrix of the features of the interest points themselves, the features with context information transmitted from the neighboring interest points, the weight matrix of the features with context information transmitted from the neighboring interest points, and the bias vector, after weighted summation processing, the transfer preference representation of the members for the interest points is obtained; Recommendation module: used to obtain the final interest point representation by summing up the members' interaction preference representations for interest points and the members' transfer preference representations for interest points, and to obtain the score of the temporary group on each candidate interest point after performing vector product operation and normalization operation on the final interest point representation and the fitting feature representation of the temporary group, and to determine the order of recommending interest points to the temporary group based on the score.

Images