Regional perception dynamic hypergraph and double-layer modeling-based interest point recommendation method

Abstract

The invention discloses an interest point recommendation method based on a region perception dynamic hypergraph and double-layer modeling, and the method comprises the steps: generating a sign-in record representation containing a region code through quadtree region coding and coarse-grained geographic information; constructing a dynamic hypergraph; capturing a dependency relationship between sign-in records in a single user target track through a self-attention mechanism, and obtaining sign-in record characterization; generating initial feature representation of each track through hypergraph volume accumulation and sign-in record representation; carrying out convolution operation on the hyperedges, fusing high-order correlation information of tracks including tracks associated with the target track in the track-track correlation matrix among the tracks, and obtaining final target track feature representation; and mapping the final target track feature representation to an ID space of a point of interest POI, and recommending a next point of interest to the user. According to the method, through regional perception and double-layer modeling, the problem of data sparsity is effectively relieved, the capability of capturing the dynamic movement mode of the user is enhanced, and the accuracy and reliability of recommendation are improved.

Description

Technical Field

[0001] The present invention relates to a point of interest recommendation method based on region-aware dynamic hypergraph and double-layer modeling, belonging to the technical field of point of interest recommendation. Background Art

[0002] With the rapid development of mobile technology and the widespread use of location-based social networks (LBSNs), users have generated a large amount of check-in data on various platforms. This data provides a rich foundation for the development of next point of interest (POI) recommendation technology, making it valuable for applications in personalized route planning, targeted advertising, and commercial site recommendations.

[0003] Existing POI recommendation methods are mainly divided into sequence-based and graph-based methods. Sequence-based methods use recurrent neural networks (RNNs) and their variants to capture sequential dependencies, but they primarily focus on short-term dependencies, limiting their ability to model diverse patterns. To overcome this limitation, self-attention networks have been introduced to model long-term dependencies, but they still primarily focus on intra-sequence relationships and ignore inter-sequence collaborative information. Furthermore, these methods rely heavily on single-user data and are susceptible to data sparsity and information cocooning problems.

[0004] Graph-based approaches represent user-POI interactions as graph structures, leveraging neighborhood information aggregation to achieve more complex global behavior modeling. However, these approaches typically only incorporate low-order POI associations, making it difficult to model the complex behaviors generated by trajectory collaboration. Hypergraph approaches transcend the limitations of binary connections and can capture higher-order relationships between POIs, users, and spatiotemporal contexts. However, existing methods still fail to fully exploit spatial associations and fail to break through the collaborative information at the individual and POI levels.

[0005] Furthermore, existing methods also have shortcomings in their utilization of geographic information. Geographic features are typically captured solely through latitude and longitude embedding. However, geographic data is sparse and nonlinear, and there is strong coupling between longitude and latitude, making single embedding ineffective. This results in the potential spatial relationships between POIs not being fully explored, thus limiting the improvement of recommendation performance.

[0006] In summary, existing methods still have shortcomings in data sparsity, information cocoon problem, high-order correlation modeling and geographic information utilization. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to overcome the defects of the existing technology and provide a point of interest recommendation method based on region-aware dynamic hypergraph and two-layer modeling. It can break through the limitations of individuals and single points and comprehensively consider the multi-dimensional association between POIs and trajectories to improve the performance and accuracy of the next POI recommendation.

[0008] Preferably, the present invention provides a method for recommending points of interest based on region-aware dynamic hypergraph and two-layer modeling, comprising:

[0009] Input the historical check-in records of the user to be predicted into the trained model, and predict the next point of interest. The historical check-in records of the user to be predicted include the user ID, the ID of the point of interest (POI), the POI category, geographic information, and a timestamp. The geographic information includes longitude and latitude.

[0010] Among them, the training completed model includes:

[0011] Step 1: Generate a check-in record representation containing the region code using quadtree region code and coarse-grained geographic information;

[0012] Step 2: Construct a dynamic hypergraph, where the dynamic hypergraph includes a check-in record-trajectory association matrix and a trajectory-trajectory association matrix;

[0013] Step 3: Use the self-attention mechanism to capture the dependencies between check-in records within a single user's target trajectory and obtain a check-in record representation. The check-in records within a single user's target trajectory include the check-in records associated with the single user's target trajectory in the check-in record-trajectory association matrix.

[0014] Step 4: Aggregate the check-in record representations through hypergraph convolution to generate the initial feature representation of each trajectory;

[0015] Step 5: Perform convolution on the hyperedge to fuse the high-order correlation information between trajectories, including the trajectories associated with the target trajectory in the trajectory-trajectory correlation matrix, to obtain the final target trajectory feature representation;

[0016] Step 6: Map the final target trajectory feature representation to the ID space of the point of interest (POI) to recommend the next point of interest to the user.

[0017] Preferably, step 1 generates a check-in record representation containing the region code using the quadtree region code and coarse-grained geographic information, including:

[0018] Get the check-in record q and the next point of interest visited by the user, q =<u,p,c,g,t> , u is the user ID, p is the ID of the point of interest POI, c is the POI category, g is the geographic information, which includes longitude and latitude, and t is the timestamp;

[0019] The latitude and longitude coordinates of the POI are divided into hierarchical regions using the quadtree region encoding method, and a base-4 quadtree string representing the quadtree region is generated;

[0020] The region index r of each check-in point is obtained by modulo the total number of grids at the final zoom level;

[0021] The region index r is used as a feature of the check-in record to generate a check-in record representation q=u,p,c,r,t> containing the region code.

[0022] Prior to step 2, a dynamic hypergraph is constructed, including:

[0023] Define a dynamic hypergraph H = (V, E), where V is the set of check-in records and E is the set of user trajectories;

[0024] Each hyperedge e∈E corresponds to the user’s trajectory divided by time Connecting tracks All sign-in records For the trajectory Middle Sign-in records;

[0025] The hypergraph structure is encoded by the node-hyperedge association matrix H1 and the hyperedge-hyperedge association matrix H2, where H1(i,j)=1 represents node v i Belongs to the hyperedge e j , H1(i,j)=0 means node v i Does not belong to the hyperedge e j ;

[0026] If and only if the trajectory s m and trajectory s n When the trajectory is the same user or similar trajectories across users, H2(m,n)=1;

[0027] Introduce edge type matrix r ε→ε Identify the connection types in the hyperedge-hyperedge incidence matrix H2, where r m,n =0 indicates intra-user association, r m,n =1 indicates collaboration among users.

[0028] Prioritize, in step 3, the self-attention mechanism is used to capture the dependencies between check-in records within a single user’s target trajectory, including:

[0029] The check-in record sequence in the target trajectory is taken as input, and the information including the ID, category, and time of each check-in record is mapped into a dense vector through the embedding layer;

[0030] The multi-head self-attention mechanism is used to calculate the dependency between POIs. The specific formula is as follows:

[0031]

[0032] Where Q(k), K(k), and V(k) are the query, key, and value vectors of the kth attention head, respectively, and d is the vector dimension;

[0033] The outputs of each attention head are concatenated and mapped back to the original dimension through linear transformation to obtain the dependency representation between check-in records in the target trajectory.

[0034] Prioritize, in step 4, check-in information is aggregated through hypergraph convolution to generate the initial feature representation of each trajectory, including:

[0035] The HypergraphTransformer layer, which uses a hypergraph convolution method, aggregates the check-in data to generate the initial representation of the trajectory. The HGTransformer includes a message assembly phase and a message propagation phase.

[0036] In the message assembly phase, the source node is hidden Edge type vector r ij , time vector t ij and the distance vector s ij Combine to get the message vector

[0037]

[0038] Where, When node j is in layer l (l) Hidden representation of r ij is the edge type vector, which is used to encode the type information of the edge from node i to node j, i.e., intra-user association or inter-user collaboration; ij is a time vector, used to encode the time information between node i and node j; ij is a distance vector, which is used to encode the distance information between node i and node j;

[0039] In the message propagation stage, the importance of each message is evaluated by multi-head scaling dot product attention, and the target node representation is updated by weighted aggregation of neighbor messages.

[0040]

[0041] Where MSDA is the multi-head scaled dot product attention function, is the message vector from node j to node i in the l-th iteration, is the set of neighbor nodes of node i.

[0042] Prioritize, in step 5, convolution operation is performed on the hyperedge to fuse high-order correlation information between trajectories to obtain the final target trajectory feature representation, including:

[0043] Perform convolution operations on hyperedges to aggregate high-order correlation information between trajectories. The specific formula is as follows:

[0044]

[0045] in, is the hidden state of the ith track in the l+1 layer, is the hidden state of the jth track in the lth layer, r ij , t ij 、s ij are the edge type, time difference and distance vector between trajectory i and trajectory j respectively;

[0046] Perform convolution operations on multiple layers of hyperedges to fuse high-order correlation information between trajectories and obtain the final target trajectory feature representation;

[0047] Based on the final target trajectory feature representation, the next point of interest is recommended to the user.

[0048] Prioritizing, convolution operations are performed on multiple layers of hyperedges to fuse high-order correlation information between trajectories to obtain the final target trajectory feature representation. Based on the final target trajectory feature representation, the next point of interest is recommended to the user, including:

[0049] Stack L-1 HG Transformer layers, each of which applies a feed-forward neural network MLP and L2 normalization;

[0050] The output of the previous layer is balanced through a linear projection and gated residual module and the output of HG Transformer

[0051]

[0052] Where l = 1, 2, ..., L-1, β is a hyperparameter representing the residual weight, The node is at the l+1th layer The intermediate hidden representation without activation function and normalization, is the weight matrix of the linear projection of the lth layer, is the corresponding bias vector; is the output of the current HG Transformer layer, which represents the hidden representation of node i after message propagation, attention mechanism and nonlinear transformation; is the hidden representation of the lth layer, Norm() is a normalization operation, and ReLU() is a commonly used activation function. is the weight matrix of the first linear transformation of this layer, is the corresponding bias vector, is the weight matrix of the second linear transformation of this layer, is the corresponding bias vector.

[0053] Prioritize, in step 4, check-in information is aggregated through hypergraph convolution to generate the initial feature representation of each trajectory, including:

[0054] Use a single-layer perceptron to represent the target trajectory Map to POIID space and predict the next location the user will visit

[0055]

[0056] Where W p and b p is the weight matrix and bias term, Softmax() is a normalization function, is the hidden representation of the target node i in the Lth layer;

[0057] The model is trained in small batches using cross entropy loss. The formula for cross entropy loss is:

[0058]

[0059] Where, is the cross entropy loss value, N is the number of samples in the current mini-batch, the sample includes the check-in record data of the users in the current mini-batch, |P| is the size of the POI set, y i is the true label of sample i, is the prediction result of the model for sample i.

[0060] Preferably, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the methods described in the first aspect when executing the program.

[0061] Preferably, the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of any one of the methods described in the first aspect when executed by a processor.

[0062] The beneficial effects achieved by the present invention are:

[0063] The present invention makes full use of geographic information and mines the potential spatial relationship between POIs through the quadtree region coding method;

[0064] This invention captures high-order collaborative information between users by constructing a hypergraph and introducing collaborative trajectories, thereby enhancing the accuracy and personalization of recommendations.

[0065] The present invention comprehensively captures the user's dynamic movement pattern through dual modeling at the POI level and the trajectory level, thereby improving the recommendation performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solution of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0067] Figure 1 Flow chart for the implementation of the method of the present invention;

[0068] Figure 2 Schematic diagram of the structure of the ReHDM framework of the method of the present invention. DETAILED DESCRIPTION

[0069] See also Figure 1 This application discloses a method for recommending points of interest based on region-aware dynamic hypergraph and two-layer modeling, including:

[0070] Input the historical check-in records of the user to be predicted into the trained model, and predict the next point of interest. The historical check-in records of the user to be predicted include the user ID, the ID of the point of interest (POI), the POI category, geographic information, and a timestamp. The geographic information includes longitude and latitude.

[0071] Among them, the training completed model includes:

[0072] Step 1: Generate a check-in record representation containing the region code using quadtree region code and coarse-grained geographic information;

[0073] Step 2: Construct a dynamic hypergraph, where the dynamic hypergraph includes a check-in record-trajectory association matrix and a trajectory-trajectory association matrix;

[0074] Step 3: Use the self-attention mechanism to capture the dependencies between check-in records within a single user's target trajectory and obtain a check-in record representation. The check-in records within a single user's target trajectory include the check-in records associated with the single user's target trajectory in the check-in record-trajectory association matrix.

[0075] Step 4: Aggregate the check-in record representations through hypergraph convolution to generate the initial feature representation of each trajectory;

[0076] Step 5: Perform convolution on the hyperedge to fuse the high-order correlation information between trajectories, including the trajectories associated with the target trajectory in the trajectory-trajectory correlation matrix, to obtain the final target trajectory feature representation;

[0077] Step 6: Map the final target trajectory feature representation to the ID space of the point of interest (POI) to recommend the next point of interest to the user.

[0078] Furthermore, in step 1, a check-in record representation containing the region code is generated using the quadtree region code and the coarse-grained geographic information, including:

[0079] Collect user identity information, user access location information, and geographical information; in this step, the present invention first collects the basic information of the user, including the user ID, the location ID visited, and the corresponding geographical information, which can be obtained from a location-based social network (LBSN) platform, such as Foursquare, Gowalla. The basic information of the collected user will be used as the basis for subsequent modeling. The check-in record q is represented as q = <u, p, c, g, t>, where u is the user ID, p is the ID of the point of interest (POI), c is the POI category, g is the geographical information, which includes longitude and latitude, and t is the timestamp;

[0080] Hierarchically divide the longitude and latitude coordinates of the POI by the quadtree region encoding method to generate a base-4 quadtree string representing the quadtree region;

[0081] By taking the modulus of the total number of grids at the final zoom level, obtain the region index r of each check-in point;

[0082] Use the region index r as a feature of the check-in record to generate a check-in record representation q = <u, p, c, r, t> containing region encoding. In step 1, with the user identity information, user access location information, and geographical information as inputs, based on the quadtree region encoding method, obtain region encoding information. In this step, the present invention uses the quadtree region encoding method to process geographical information. Specifically, the present invention divides the earth's surface into grids at different levels, and each grid is represented by a unique quadtree code. In this way, geographical coordinates can be converted into discrete region encodings, which is convenient for subsequent hypergraph construction. The specific formula is as follows:

[0083] Region encoding = Quadkey-based Regional Encoding(Geographical coordinates, Zoom level)

[0084] Among them, Quadkey-based Regional Encoding represents the quadtree region encoding function, the geographical coordinates are the input longitude and latitude information, and the zoom level represents the encoding accuracy.

[0085] Furthermore, in step 2, construct a dynamic hypergraph, including:

[0086] Define the dynamic hypergraph H = (V, E), where V is the set of check-in records and E is the set of user trajectories;

[0087] Each hyperedge e ∈ E corresponds to the trajectory of the user divided by time Connect the trajectories All the check-in records within For the trajectory Middle Sign-in records;

[0088] The hypergraph structure is encoded by the node-hyperedge association matrix H1 and the hyperedge-hyperedge association matrix H2, where H1(i,j)=1 represents node v i Belongs to the hyperedge e j , H1(i,j)=0 means node v i Does not belong to the hyperedge e j ;

[0089] If and only if the trajectory s m and trajectory s n When the trajectory is the same user or similar trajectories across users, H2(m,n)=1;

[0090] Introduce edge type matrix r ε→ε Identify the connection types in the hyperedge-hyperedge incidence matrix H2, where r m,n =0 indicates intra-user association, r m,n =1 indicates collaboration among users.

[0091] Furthermore, in step 3, the self-attention mechanism is used to capture the dependencies between check-in records within a single user's target trajectory, including:

[0092] The check-in sequence in the target trajectory is taken as input, and the information including the ID, category, and time of each check-in record is mapped into a dense vector through the embedding layer;

[0093] The multi-head self-attention mechanism is used to calculate the dependency between check-in records. The specific formula is as follows:

[0094]

[0095] Where Q(k), K(k), and V(k) are the query, key, and value vectors of the kth attention head, respectively, and d is the vector dimension;

[0096] The outputs of each attention head are concatenated and mapped back to the original dimension through linear transformation to obtain the dependency representation between check-in records in the target trajectory.

[0097] Furthermore, in step 4, the check-in information is aggregated through the hypergraph convolution method to generate the initial feature representation of each trajectory, including:

[0098] The HypergraphTransformer layer, which uses a hypergraph convolution method, aggregates the check-in data to generate the initial representation of the trajectory. The HGTransformer includes a message assembly phase and a message propagation phase.

[0099] In the message assembly phase, the source node is hidden Edge type vector rij , time vector t ij and the distance vector s ij Combine to get the message vector

[0100]

[0101] Where, When node j is in layer l (l) Hidden representation of r ij is the edge type vector, which is used to encode the type information of the edge from node i to node j, i.e., intra-user association or inter-user collaboration; ij is a time vector, used to encode the time information between node i and node j; ij is a distance vector, which is used to encode the distance information between node i and node j;

[0102] In the message propagation stage, the importance of each message is evaluated by multi-head scaling dot product attention, and the target node representation is updated by weighted aggregation of neighbor messages.

[0103]

[0104] Where MSDA is the multi-head scaled dot product attention function, is the message vector from node j to node i in the l-th iteration, is the set of neighbor nodes of node i.

[0105] Furthermore, in step 5, a convolution operation is performed on the hyperedge to fuse the high-order correlation information between trajectories to obtain the final target trajectory feature representation, including:

[0106] Perform convolution operations on hyperedges to aggregate high-order correlation information between trajectories. The specific formula is as follows:

[0107]

[0108] in, is the hidden state of the ith track in the l+1 layer, is the hidden state of the jth track in the lth layer, r ij , t ij 、s ij are the edge type, time difference and distance vector between trajectory i and trajectory j respectively;

[0109] Perform convolution operations on multiple layers of hyperedges to fuse high-order correlation information between trajectories and obtain the final target trajectory feature representation;

[0110] Based on the final target trajectory feature representation, the next point of interest is recommended to the user.

[0111] Furthermore, convolution operations are performed on multiple layers of hyperedges to fuse high-order correlation information between trajectories to obtain the final target trajectory feature representation. Based on the final target trajectory feature representation, the next point of interest is recommended to the user, including:

[0112] Stack L-1 HG Transformer layers, each of which applies a feedforward neural network (MLP) and L2 normalization. MLP is a feedforward neural network consisting of multiple fully connected layers and nonlinear activation functions. L2 normalization is an operation that normalizes vectors to unit length.

[0113] The output of the previous layer is balanced through a linear projection and gated residual module and the output of HG Transformer Effectively integrates knowledge of collaborative trajectories:

[0114]

[0115] Where l = 1, 2, ..., L-1, β is a hyperparameter representing the residual weight, The node is at the l+1th layer The intermediate hidden representation without activation function and normalization, is the weight matrix of the linear projection of the lth layer, is the corresponding bias vector; is the output of the current HG Transformer layer, which represents the hidden representation of node i after message propagation, attention mechanism and nonlinear transformation; is the hidden representation of the lth layer, Norm() is a normalization operation, and ReLU() is a commonly used activation function. is the weight matrix of the first linear transformation of this layer, is the corresponding bias vector, is the weight matrix of the second linear transformation of this layer, is the corresponding bias vector.

[0116] Furthermore, in step 4, the check-in information is aggregated through hypergraph convolution to generate the initial feature representation of each trajectory, including:

[0117] Use a single-layer perceptron to represent the target trajectory Map to POIID space and predict the next location the user will visit

[0118]

[0119] Where W p and bp is the weight matrix and bias term, Softmax() is a normalization function, is the hidden representation of the target node i in the Lth layer;

[0120] The model is trained in mini-batch using cross entropy loss. The model is the entire network including all operations from step 1 to step 6 above:

[0121]

[0122] Where, is the cross entropy loss value, N is the number of samples in the current mini-batch, and the samples include the check-in records of the users in the current mini-batch. Each check-in record q contains the following parameters<u,p,c,g,t> , |P| is the size of the POI set, y i is the true label of sample i, is the prediction result of the model for sample i. The model is the entire network including all operations from step 1 to step 6 above.

[0123] In an embodiment of the present application, the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above methods when executing the program.

[0124] In an embodiment of the present application, the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of any of the above methods when executed by a processor.

[0125] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0126] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention as disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not invented herein, and the description and examples are to be considered merely as exemplary.

[0127] The above specific implementation methods further illustrate the purpose, technical solutions and beneficial effects of this application in detail. It should be understood that the above are only specific implementation methods of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solutions of this application should be included in the scope of protection of this application.

Claims

1. A method for recommending points of interest based on region-aware dynamic hypergraph and two-layer modeling, characterized in that: include: Input the historical check-in records of the user to be predicted into the trained model, and predict the next point of interest. The historical check-in records of the user to be predicted include the user ID, the ID of the point of interest (POI), the POI category, geographic information, and a timestamp. The geographic information includes longitude and latitude. Among them, the training completed model includes: Step 1: Generate a check-in record representation containing the region code using quadtree region code and coarse-grained geographic information; Step 2: Construct a dynamic hypergraph, where the dynamic hypergraph includes a check-in record-trajectory association matrix and a trajectory-trajectory association matrix; Step 3: Use the self-attention mechanism to capture the dependencies between check-in records within a single user's target trajectory and obtain a check-in record representation. The check-in records within a single user's target trajectory include the check-in records associated with the single user's target trajectory in the check-in record-trajectory association matrix. Step 4: Aggregate the check-in record representations through hypergraph convolution to generate the initial feature representation of each trajectory; Step 5: Perform convolution on the hyperedge to fuse the high-order correlation information between trajectories, including the trajectories associated with the target trajectory in the trajectory-trajectory correlation matrix, to obtain the final target trajectory feature representation; Step 6: Map the final target trajectory feature representation to the ID space of the point of interest (POI) to recommend the next point of interest to the user.

2. The method for recommending points of interest based on region-aware dynamic hypergraph and two-layer modeling according to claim 1, characterized in that: Step 1: Generate a check-in record representation containing the region code using the quadtree region code and coarse-grained geographic information, including: Get the check-in record q and the next point of interest visited by the user, q =<u,p,c,g,t> , u is the user ID, p is the ID of the point of interest POI, c is the POI category, g is the geographic information, which includes longitude and latitude, and t is the timestamp; The latitude and longitude coordinates of the POI are divided into hierarchical regions using the quadtree region encoding method, and a base-4 quadtree string representing the quadtree region is generated; The region index r of each check-in point is obtained by modulo the total number of grids at the final zoom level; The region index r is used as a feature of the check-in record to generate a check-in record representation q containing the region code =<u,p,c,r,t> .

3. The method for recommending points of interest based on region-aware dynamic hypergraph and two-layer modeling according to claim 1, characterized in that: In step 2, a dynamic hypergraph is constructed, including: Define a dynamic hypergraph H = (V, E), where V is the set of check-in records and E is the set of user trajectories; Each hyperedge e∈E corresponds to the user’s trajectory divided by time Connecting tracks All sign-in records For the trajectory Middle Sign-in records; The hypergraph structure is encoded by the node-hyperedge association matrix H1 and the hyperedge-hyperedge association matrix H2, where H1(i,j)=1 represents node v i Belongs to the hyperedge e j , H1(i,j)=0 means node v i Does not belong to the hyperedge e j ; If and only if the trajectory s m and trajectory s n When the trajectory is the same user or similar trajectories across users, H2(m,n)=1; Introduce edge type matrix r ε→ε Identify the connection types in the hyperedge-hyperedge incidence matrix H2, where r m,n =0 indicates intra-user association, r m,n =1 indicates collaboration among users.

4. The method for recommending points of interest based on region-aware dynamic hypergraph and two-layer modeling according to claim 1, characterized in that: In step 3, the self-attention mechanism is used to capture the dependencies between check-in records within a single user’s target trajectory, including: The check-in record sequence in the target trajectory is taken as input, and the information including the ID, category, and time of each check-in record is mapped into a dense vector through the embedding layer; The multi-head self-attention mechanism is used to calculate the dependency between POIs. The specific formula is as follows: Where Q(k), K(k), and V(k) are the query, key, and value vectors of the kth attention head, respectively, and d is the vector dimension; The outputs of each attention head are concatenated and mapped back to the original dimension through linear transformation to obtain the dependency representation between check-in records in the target trajectory.

5. The method for recommending points of interest based on region-aware dynamic hypergraph and two-layer modeling according to claim 1, characterized in that: In step 4, the check-in information is aggregated through the hypergraph convolution method to generate the initial feature representation of each trajectory, including: The HypergraphTransformer layer, which uses a hypergraph convolution method, aggregates the check-in data to generate the initial representation of the trajectory. The HGTransformer includes a message assembly phase and a message propagation phase. In the message assembly phase, the source node is hidden Edge type vector r ij , time vector t ij and the distance vector s ij Combine to get the message vector Where, When node j is in layer l (l) Hidden representation of r ij is the edge type vector, which is used to encode the type information of the edge from node i to node j, i.e., intra-user association or inter-user collaboration; ij is a time vector, used to encode the time information between node i and node j; ij is a distance vector, which is used to encode the distance information between node i and node j; In the message propagation stage, the importance of each message is evaluated by multi-head scaling dot product attention, and the target node representation is updated by weighted aggregation of neighbor messages. Where MSDA is the multi-head scaled dot product attention function, is the message vector from node j to node i in the l-th iteration, is the set of neighbor nodes of node i.

6. The method for recommending points of interest based on region-aware dynamic hypergraph and two-layer modeling according to claim 1, characterized in that: In step 5, a convolution operation is performed on the hyperedge to fuse the high-order correlation information between the trajectories and obtain the final target trajectory feature representation, including: Perform convolution operations on hyperedges to aggregate high-order correlation information between trajectories. The specific formula is as follows: in, is the hidden state of the i-th track in the l+1 layer, is the hidden state of the jth trajectory in the lth layer, r ij , t ij 、s ij are the edge type, time difference and distance vector between trajectory i and trajectory j respectively; Perform convolution operations on multiple layers of hyperedges to fuse high-order correlation information between trajectories and obtain the final target trajectory feature representation; Based on the final target trajectory feature representation, the next point of interest is recommended to the user.

7. The method for recommending points of interest based on region-aware dynamic hypergraph and two-layer modeling according to claim 6, characterized in that: Perform convolution operations on multiple layers of hyperedges to fuse high-order correlation information between trajectories and obtain the final target trajectory feature representation; Based on the final target trajectory feature representation, the next point of interest is recommended to the user, including: Stack L-1 HG Transformer layers, each of which applies a feed-forward neural network MLP and L2 normalization; The output of the previous layer is balanced through a linear projection and gated residual module and the output of HG Transformer Where l = 1, 2, ..., L-1, β is a hyperparameter representing the residual weight, For the l+1th layer node The intermediate hidden representation without activation function and normalization, is the weight matrix of the linear projection of the lth layer, is the corresponding bias vector; is the output of the current HG Transformer layer, which represents the hidden representation of node i after message propagation, attention mechanism and nonlinear transformation; is the hidden representation of the previous layer (layer l), Norm() is a normalization operation, and ReLU() is a commonly used activation function. is the weight matrix of the first linear transformation of this layer, is the corresponding bias vector, is the weight matrix of the second linear transformation of this layer, is the corresponding bias vector.

8. The method for recommending points of interest based on region-aware dynamic hypergraph and two-layer modeling according to claim 1, characterized in that: In step 4, the check-in information is aggregated through hypergraph convolution to generate the initial feature representation of each trajectory, including: Use a single-layer perceptron to represent the target trajectory Map to POI ID space and predict the next location the user will visit Where W p and b p is the weight matrix and bias term, Softmax() is a normalization function, is the hidden representation of the target node i in the Lth layer; The model is trained in small batches using cross entropy loss. The formula for cross entropy loss is: Where, is the cross entropy loss value, N is the number of samples in the current mini-batch, the sample includes the check-in record data of the users in the current mini-batch, |P| is the size of the POI set, y i is the true label of sample i, is the prediction result of the model for sample i.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images