A point of interest recommendation method based on regional perception dynamic hypergraph and double-layer modeling

Abstract

The application discloses a point of interest recommendation method based on regional perception dynamic hypergraph and double-layer modeling, generates a check-in record representation containing a region code through a quadtree region code and coarse-grained geographic information; constructs a dynamic hypergraph; captures the dependency relationship between the check-in records in a single user target track through a self-attention mechanism to obtain check-in record representation; aggregates the check-in record representation through hypergraph convolution to generate an initial feature representation of each track; performs convolution operation on a hyperedge to fuse high-order correlation information of the tracks associated with the target track in a track-track correlation matrix between the tracks to obtain a final target track feature representation; and maps the final target track feature representation to an ID space of a point of interest (POI) to recommend a next point of interest for a user. The application effectively alleviates the data sparsity problem, enhances the capture ability of the dynamic moving mode of the user, and improves the accuracy and reliability of the recommendation through regional perception and double-layer modeling.

Description

Technical Field

[0001] This invention relates to a method for recommending points of interest based on region-aware dynamic hypergraphs and two-layer modeling, belonging to the field of point of interest recommendation technology. Background Technology

[0002] With the rapid development of mobile technology and the widespread application 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 Points of Interest (POI) recommendation technology, making it of significant application value in areas such as personalized route planning, precise advertising, and commercial site selection recommendations.

[0003] Existing POI recommendation methods are mainly divided into sequence-based methods and graph-based methods. Sequence-based methods capture sequence dependencies using recurrent neural networks (RNNs) and their variants, 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 mainly focus on intra-sequence relationships, neglecting collaborative information between sequences. Furthermore, these methods heavily rely on single-user data and are susceptible to data sparsity and information cocoon problems.

[0004] Graph-based methods represent user-POI interactions as graph structures, leveraging neighborhood information aggregation to model more complex global behaviors. However, these methods typically only include low-order POI associations, making it difficult to model complex behaviors arising from trajectory collaboration. Hypergraph methods overcome the limitations of binary connections, capturing higher-order relationships between POIs, users, and spatiotemporal contexts; however, existing methods still fail to fully explore spatial associations and do not break through the collaborative information at the individual and POI levels.

[0005] Furthermore, existing methods have limitations in utilizing geographic information. They typically capture geographic features solely through latitude and longitude embedding, but geographic data is sparse and nonlinear, with strong coupling between latitude and longitude coordinates. Sole embedding yields poor results, which prevents the full exploitation of potential spatial relationships between POIs, thus limiting the improvement of recommendation performance.

[0006] In summary, existing methods still have shortcomings in areas such as data sparsity, information cocoon problem, high-order association 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 prior art and provide a point of interest recommendation method based on region-aware dynamic hypergraph and two-layer modeling. This method can break through the limitations of individuals and single points and comprehensively consider the multi-dimensional correlation between POI and trajectory, so as to improve the performance and accuracy of the next POI recommendation.

[0008] Prior to this invention, an interest point recommendation method based on region-aware dynamic hypergraph and two-layer modeling is provided, comprising:

[0009] The historical check-in records of the user to be predicted are input into the trained model, and the prediction output is obtained to obtain 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, geographical information, and timestamp. The geographical information includes longitude and latitude.

[0010] The trained models include:

[0011] Step 1: Generate a sign-in record representation containing regional codes using quadtree regional coding and coarse-grained geographic information;

[0012] Step 2, construct a dynamic hypergraph, which includes the sign-in record-trajectory association matrix and the trajectory-trajectory association matrix;

[0013] Step 3: Capture the dependencies between check-in records within a single user's target trajectory using a self-attention mechanism to 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 sign-in record representations through hypergraph convolution to generate the initial feature representation of each trajectory;

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

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

[0017] Prioritizes step 1, generating a check-in record representation containing regional codes using quadtree region coding and coarse-grained geographic information, including:

[0018] Get sign-in records And the user's next point of interest. , For user ID, For the ID of the Point of Interest (POI), For POI category, Geographic information includes longitude and latitude. For timestamps;

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

[0020] The region index of each check-in point is obtained by taking the modulo of the final scaling level and the total number of grids. ;

[0021] regional index As a feature of the check-in record, a check-in record representation containing a region code is generated. .

[0022] Prioritizes step 2, which involves constructing a dynamic hypergraph, including:

[0023] Define a dynamic hypergraph ,in For the collection of sign-in records, A collection of user trajectories;

[0024] Each super edge Corresponding user trajectories divided by time , connecting trajectory All check-in records , For trajectory The Middle One sign-in record;

[0025] Through the node-hyperedge correlation matrix Hyperedge-Hyperedge Incidence Matrix Encoding hypergraph structure, in which Represents a node Belongs to superedge , 0 represents a node Not a hyperedge ;

[0026] If and only if the trajectory and trajectory When the trajectory is the same user trajectory or a similar trajectory across users, ;

[0027] Introducing the edge type matrix Identifying Hyperedge-Hyperedge Incidence Matrix The connection types in the data, where Indicates user-internal association. This indicates collaboration between users.

[0028] Prioritizing 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 sequence of check-in records within 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 dependencies between Points of Interest (POIs), and the specific formula is as follows:

[0031]

[0032] in, The first A query, key, and value vector for each attention head. For vector dimensions;

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

[0034] Prioritizes step 4 by aggregating check-in information using hypergraph convolution to generate an initial feature representation for each trajectory, including:

[0035] Using the hypergraph convolution method Layered aggregation of check-in data generates an initial trajectory representation, where It includes the message assembly phase and the message propagation phase;

[0036] During the message assembly phase, the source node representation is hidden. Edge type vector Time vector and distance vector Combine to obtain the message vector :

[0037] ,

[0038] In the formula, When the node is at layer l Hidden representations; This is an edge type vector used to encode the type information of the edge from node i to node j, i.e., intra-user association or inter-user collaboration; This is a time vector used to encode the time information between node i and node j; This is a distance vector used to encode the distance information between node i and node j;

[0039] During the message propagation phase, the importance of each message is evaluated through multi-head scaled dot product attention, and the neighbor messages are weighted and aggregated to update the target node representation. :

[0040] ,

[0041] In the formula, MSDA is the multi-head scaled dot product attention function. Let i be the message vector from node j to node i in the l-th layer iteration. Let i be the set of neighboring nodes of node i.

[0042] Prioritizes step 5 by performing convolution operations on the hyperedges to fuse high-order correlation information between trajectories, thereby obtaining the final target trajectory feature representation, including:

[0043] The hyperedge is convolved to aggregate higher-order correlation information between trajectories. The specific formula is as follows:

[0044] ,

[0045] in, For the first The trajectory at the th The hidden state of the layer For the first The hidden state of a trajectory in layer l , , Trajectories and trajectory The edge type, time difference, and distance vector between them;

[0046] Convolution operations are performed on the multi-layer hyperedges to fuse the 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] Firstly, convolution operations are performed on multiple hyperedges to fuse high-order correlation information between trajectories, resulting in the final target trajectory feature representation. Based on this final target trajectory feature representation, the next point of interest is recommended to the user, including:

[0049] Stack L-1 Layer, each The layer applies a feedforward neural network and Normalization;

[0050] The output of the previous layer is balanced using a linear projection and a gated residual module. and Output :

[0051]

[0052]

[0053] in , It is a hyperparameter representing the residual weight. When the node is at layer l+1 The intermediate hidden representations that have not undergone activation functions and normalization, Let be the weight matrix of the linear projection of the l-th layer. This is the corresponding bias vector; For the present The output of the layer represents the hidden representation of node i after message propagation, attention mechanism and nonlinear transformation; For the hidden representation of the l-th layer, () is a normalization operation. () is a commonly used activation function. This is the weight matrix for the first linear transformation of this layer. For the corresponding bias vector, This is the weight matrix for the second linear transformation of this layer. This is the corresponding bias vector.

[0054] Prioritizes step 4, where check-in information is aggregated via hypergraph convolution to generate an initial feature representation for each trajectory, including:

[0055] Representing the target trajectory using a single-layer perceptron Mapped to Space, predicting the user's next location. :

[0056]

[0057] in and These are the weight matrix and the bias terms. () is a normalization function. Let i be the hidden representation of the target node i in the Lth layer;

[0058] The model is trained using cross-entropy loss in mini-batch training. The formula for cross-entropy loss is:

[0059] ,

[0060] In the formula, Here, N represents the cross-entropy loss value, and N is the number of samples in the current mini-batch, which includes the check-in records of users in the current mini-batch. The size of the POI set. For the true label of sample i, This represents the model's prediction result for sample i.

[0061] Preferably, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described in any of the first aspects.

[0062] Preferably, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any one of the first aspects.

[0063] The beneficial effects achieved by this invention are as follows:

[0064] This invention utilizes a quadtree region coding method to fully leverage geographic information and uncover potential spatial relationships between Points of Interest (POIs).

[0065] This invention enhances the accuracy and personalization of recommendations by constructing a hypergraph and introducing collaborative trajectories to capture high-order collaborative information among users.

[0066] This invention comprehensively captures users' dynamic movement patterns through dual modeling at the POI level and trajectory level, thereby improving recommendation performance. Attached Figure Description

[0067] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention;

[0069] Figure 2 This is a schematic diagram of the ReHDM framework of the method of the present invention. Detailed Implementation

[0070] See Figure 1 This application discloses an interest point recommendation method based on region-aware dynamic hypergraph and two-layer modeling, including:

[0071] The historical check-in records of the user to be predicted are input into the trained model, and the prediction output is obtained to obtain 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, geographical information, and timestamp. The geographical information includes longitude and latitude.

[0072] The trained models include:

[0073] Step 1: Generate a sign-in record representation containing regional codes using quadtree regional coding and coarse-grained geographic information;

[0074] Step 2, construct a dynamic hypergraph, which includes the sign-in record-trajectory association matrix and the trajectory-trajectory association matrix;

[0075] Step 3: Capture the dependencies between check-in records within a single user's target trajectory using a self-attention mechanism to 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.

[0076] Step 4: Aggregate the sign-in record representations through hypergraph convolution to generate the initial feature representation of each trajectory;

[0077] Step 5: Perform convolution operation on the hyperedge to fuse the high-order correlation information between trajectories, including the trajectory-trajectory correlation matrix, which is associated with the target trajectory, to obtain the final target trajectory feature representation;

[0078] Step 6: Map the final target trajectory feature representation to the ID space of Points of Interest (POIs) to recommend the next POI to the user.

[0079] Further, in step 1, a check-in record representation containing a region code is generated using quadtree region coding and coarse-grained geographic information, including:

[0080] The process involves collecting user identity information, user access location information, and geographic information. In this step, the invention first collects basic user information, including user ID, visited location IDs, and corresponding geographic information. This data can be obtained from location-based social network (LBSN) platforms such as Foursquare and Gowalla. The collected basic user information will serve as the basis for subsequent modeling. Check-in records will be included. Represented as ,in For user ID, For the ID of the Point of Interest (POI), For POI category, Geographic information includes longitude and latitude. For timestamps;

[0081] The latitude and longitude coordinates of the POI are hierarchically divided into regions using the quadtree region encoding method, generating a base-4 quadtree string representing the quadtree region;

[0082] The region index of each check-in point is obtained by taking the modulo of the final scaling level and the total number of grids. ;

[0083] regional index As a feature of the check-in record, a check-in record representation containing a region code is generated. In step 1, the user identity information, user access location information, and geographic information are used as inputs. Based on the quadtree region coding method, region coding information is obtained. In this step, the present invention uses the quadtree region coding method to process geographic information. Specifically, the present invention divides the Earth's surface into grids of different levels, with each grid represented by a unique quadtree code. In this way, geographic coordinates can be converted into discrete region codes, facilitating subsequent hypermap construction. The specific formula is as follows:

[0084] Area Code (Geographic coordinates, zoom level)

[0085] in, This represents a quadtree region coding function, where the geographic coordinates are the latitude and longitude information of the input, and the scaling level represents the coding precision.

[0086] Furthermore, in step 2, constructing a dynamic hypergraph includes:

[0087] Define a dynamic hypergraph ,in For the collection of sign-in records, A collection of user trajectories;

[0088] Each super edge Corresponding user trajectories divided by time , connecting trajectory All check-in records , For trajectory The Middle One sign-in record;

[0089] Through the node-hyperedge correlation matrix Hyperedge-Hyperedge Incidence Matrix Encoding hypergraph structure, in which Represents a node Belongs to superedge , 0 represents a node Not a hyperedge ;

[0090] If and only if the trajectory and trajectory When the trajectory is the same user trajectory or a similar trajectory across users, ;

[0091] Introducing the edge type matrix Identifying Hyperedge-Hyperedge Incidence Matrix The connection types in the data, where Indicates user-internal association. This indicates collaboration between users.

[0092] Furthermore, in step 3, a self-attention mechanism is used to capture the dependencies between check-in records within a single user's target trajectory, including:

[0093] The check-in sequence within 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;

[0094] The multi-head self-attention mechanism is used to calculate the dependencies between check-in records. The specific formula is as follows:

[0095]

[0096] in, The first A query, key, and value vector for each attention head. For vector dimensions;

[0097] The outputs of each attention head are concatenated and mapped back to the original dimension through a linear transformation to obtain a representation of the dependency relationship between check-in records within the target trajectory.

[0098] Furthermore, in step 4, the check-in information is aggregated using hypergraph convolution to generate an initial feature representation for each trajectory, including:

[0099] Using the hypergraph convolution method Layered aggregation of check-in data generates an initial trajectory representation, where It includes the message assembly phase and the message propagation phase;

[0100] During the message assembly phase, the source node representation is hidden. Edge type vector Time vector and distance vector Combine to obtain the message vector :

[0101] ,

[0102] In the formula, When the node is at layer l Hidden representations; This is an edge type vector used to encode the type information of the edge from node i to node j, i.e., intra-user association or inter-user collaboration; This is a time vector used to encode the time information between node i and node j; This is a distance vector used to encode the distance information between node i and node j;

[0103] During the message propagation phase, the importance of each message is evaluated through multi-head scaled dot product attention, and the neighbor messages are weighted and aggregated to update the target node representation. :

[0104] ,

[0105] In the formula, MSDA is the multi-head scaled dot product attention function. Let i be the message vector from node j to node i in the l-th iteration. Let i be the set of neighboring nodes of node i.

[0106] Furthermore, in step 5, a convolution operation is performed on the hyperedges to fuse high-order correlation information between trajectories, resulting in the final target trajectory feature representation, including:

[0107] The hyperedge is convolved to aggregate higher-order correlation information between trajectories. The specific formula is as follows:

[0108] ,

[0109] in, For the first The trajectory at the th The hidden state of the layer For the first The hidden state of a trajectory in layer l , , Trajectories and trajectory The edge type, time difference, and distance vector between them;

[0110] Convolution operations are performed on the multi-layer hyperedges to fuse the high-order correlation information between trajectories and obtain the final target trajectory feature representation;

[0111] Based on the final target trajectory feature representation, the next point of interest is recommended to the user.

[0112] Furthermore, convolution operations are performed on the multi-layer hyperedges to fuse high-order correlation information between trajectories, resulting in the final target trajectory feature representation. Based on this final target trajectory feature representation, the next point of interest is recommended to the user, including:

[0113] Stack L-1 Layer, each The layer applies a feedforward neural network and Normalization, MLP is a type of feedforward neural network consisting of multiple fully connected layers and non-linear activation functions. L2 normalization is an operation that normalizes a vector to a unit length.

[0114] The output of the previous layer is balanced using a linear projection and a gated residual module. and Output This effectively integrates knowledge of collaborative trajectories:

[0115]

[0116]

[0117] in , It is a hyperparameter representing the residual weight. When the node is at layer l+1 The intermediate hidden representations that have not undergone activation functions and normalization, Let be the weight matrix of the linear projection of the l-th layer. This is the corresponding bias vector; For the present The output of the layer represents the hidden representation of node i after message propagation, attention mechanism and nonlinear transformation; For the hidden representation of the l-th layer, () is a normalization operation. () is a commonly used activation function. This is the weight matrix for the first linear transformation of this layer. For the corresponding bias vector, This is the weight matrix for the second linear transformation of this layer. This is the corresponding bias vector.

[0118] Further, in step 4, the check-in information is aggregated through hypergraph convolution to generate an initial feature representation for each trajectory, including:

[0119] Representing the target trajectory using a single-layer perceptron Mapped to Space, predicting the user's next location. :

[0120]

[0121] in and These are the weight matrix and the bias terms. () is a normalization function. Let i be the hidden representation of the target node i in the Lth layer;

[0122] The model is trained in mini-batch using cross-entropy loss. The model is the entire network that includes all operations from steps 1 to 6 above.

[0123] ,

[0124] In the formula, The cross-entropy loss value is given by N, where N is the number of samples in the current mini-batch. The samples include the check-in records of users in the current mini-batch, and each check-in record... Includes the following parameters , The size of the POI set. For the true label of sample i, The model is the prediction result of the model for sample i. The model is the entire network that includes all the operations from steps 1 to 6 above.

[0125] In this embodiment of the application, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described above.

[0126] In this application embodiment, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0127] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0128] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention described herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not invented herein. The specification and embodiments are to be considered exemplary only.

[0129] The above specific embodiments further illustrate the purpose, technical solution and beneficial effects of this application. It should be understood that the above are only specific embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of this application should be included within the scope of protection of this application.

Claims

1. An interest point recommendation method based on region-aware dynamic hypergraph and two-layer modeling, characterized in that, include: The historical check-in records of the user to be predicted are input into the trained model, and the prediction output is obtained to obtain 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, geographical information, and timestamp. The geographical information includes longitude and latitude. The trained models include: Step 1: Generate a sign-in record representation containing regional codes using quadtree regional coding and coarse-grained geographic information; Step 2, construct a dynamic hypergraph, which includes the sign-in record-trajectory association matrix and the trajectory-trajectory association matrix; Step 3: Capture the dependencies between check-in records within a single user's target trajectory using a self-attention mechanism to 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 sign-in record representations through hypergraph convolution to generate the initial feature representation of each trajectory; Step 5: Perform a convolution operation on the hyperedge to fuse the high-order correlation information of the trajectory associated with the target trajectory in the trajectory-trajectory correlation matrix, and obtain the final target trajectory feature representation; the hyperedge corresponds to the user's trajectory divided by time and connects all check-in records within the trajectory; Step 6: Map the final target trajectory feature representation to the ID space of Points of Interest (POIs) to recommend the next POI to the user.

2. The interest point recommendation method based on region-aware dynamic hypergraph and two-layer modeling as described in claim 1, characterized in that, Step 1: Generate a check-in record representation containing the region code using quadtree region coding and coarse-grained geographic information, including: Get sign-in records And the user's next point of interest. , For user ID, For the ID of the Point of Interest (POI), For POI category, Geographic information, which includes longitude and latitude, For timestamps; The latitude and longitude coordinates of the POI are hierarchically divided into regions using the quadtree region encoding method, generating a base-4 quadtree string representing the quadtree region; The region index of each check-in point is obtained by taking the modulo of the total number of grids at the final scaling level; By using the regional index as a feature of the check-in record, a check-in record representation containing the regional code is generated.

3. The interest point recommendation method based on region-aware dynamic hypergraph and two-layer modeling according to claim 1, characterized in that, Step 2 involves constructing a dynamic hypergraph, including: Define a dynamic hypergraph ,in For the collection of sign-in records, A collection of user trajectories; Each super edge Corresponding user trajectories divided by time , connecting trajectory All check-in records , For trajectory The Middle One sign-in record; Through the node-hyperedge correlation matrix Hyperedge-Hyperedge Incidence Matrix Encoding hypergraph structure, in which Represents a node Belongs to superedge , 0 represents a node Not a hyperedge ; If and only if the trajectory and trajectory When the trajectory is the same user trajectory or a similar trajectory across users, .

4. The interest point recommendation method based on region-aware dynamic hypergraph and two-layer modeling as described in 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 sequence of check-in records within 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 dependencies between Points of Interest (POIs), and the specific formula is as follows: ； Among them, Q (k) K (k) V (k) The first A query, key, and value vector for each attention head. For vector dimensions; The outputs of each attention head are concatenated and mapped back to the original dimension through a linear transformation to obtain a representation of the dependency relationship between check-in records within the target trajectory.

5. The interest point recommendation method 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 using hypergraph convolution to generate an initial feature representation for each trajectory, including: Using the hypergraph convolution method Layered aggregation of check-in data generates an initial trajectory representation, where It includes the message assembly phase and the message propagation phase; During the message assembly phase, the first The source node of the layer Hidden Representations Edge type vector Time vector and distance vector Combine to obtain from the source node Transmit to the target node message vector : ; In the formula, When the node is at layer l Hidden representations; This is an edge type vector used to encode the type information of the edge from node i to node j, i.e., intra-user association or inter-user collaboration; This is a time vector used to encode the time information between node i and node j; This is a distance vector used to encode the distance information between node i and node j; During the message propagation phase, the importance of each message is evaluated through multi-head scaled dot product attention, neighbor messages are weighted and aggregated, and the target node is updated. The representation of: ； In the formula, MSDA is the multi-head scaled dot product attention function; For the first Layer target node Hidden representations; For the first Layer target node Hidden representations; For nodes The set of neighboring nodes.

6. The interest point recommendation method 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 high-order correlation information between trajectories, resulting in the final target trajectory feature representation, including: The hyperedge is convolved to aggregate higher-order correlation information between trajectories. The specific formula is as follows: ； in, For the first The trajectory at the th Layer The output representation obtained after aggregating neighbor trajectory information; For the first The trajectory at the th Hidden representations of layers; For the first The neighbor trajectory in the first Hidden representations of layers; , , Trajectories With trajectory The edge type vector, time difference vector, and distance vector between them; Convolution operations are performed on multiple superedges to fuse high-order correlation information between trajectories, resulting in the final target trajectory feature representation.

7. The interest point recommendation method based on region-aware dynamic hypergraph and two-layer modeling according to claim 6, characterized in that, Convolutional operations are performed on multiple hyperedges to fuse high-order correlation information between trajectories, resulting in the final target trajectory feature representation; including: Stack L-1 Layer, each The layer applies a feedforward neural network and Normalization; The output of the previous layer is balanced using a linear projection and a gated residual module. and Output : ； ； in , It is a hyperparameter representing the residual weight. This represents the intermediate hidden representation of the i-th trajectory at layer l+1, before activation and normalization. Let be the weight matrix of the linear projection of the l-th layer. This is the corresponding bias vector; For the present The output of the layer represents the hidden representation of the i-th trajectory after message propagation, attention mechanism and nonlinear transformation; For the hidden representation of the (l+1)th layer, () is a normalization operation. () is a commonly used activation function. Let be the weight matrix of the first linear transformation of the l-th layer. For the corresponding bias vector, Let be the weight matrix of the second linear transformation of the l-th layer. This is the corresponding bias vector.

8. The interest point recommendation method based on region-aware dynamic hypergraph and two-layer modeling according to claim 1, characterized in that, Representing the target trajectory using a single-layer perceptron Mapped to Space, predicting the user's next location. : ； in and These are the weight matrix and the bias terms. () is a normalization function. For target node i in the th Hidden representations of layers; Cross-entropy loss is used to train the model in mini-batch.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 8.

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

Images