A multi-view interest point recommendation method based on region modeling and boundary optimization

Abstract

The application provides a multi-view point of interest recommendation method based on regional modeling and boundary optimization, and belongs to the technical field of recommendation systems and graph neural networks; first, a plurality of high-order hypergraph structures of spatial view, time view and collaborative view are constructed, and a multi-scale space modeling method driven by a center node is used to capture regional-level spatial dependency; subsequently, hypergraph convolution is used to learn high-order structure representation under different views, and a hierarchical progressive fusion mechanism is used to realize multi-view semantic alignment; cross-view contrast learning is used to enhance the representation consistency between different views, a conditional diffusion model is used to generate difficult negative samples near the boundary, and a dynamic ranking margin reinforcement model is used to enhance the discrimination ability of similar points of interest; finally, accurate prediction of the next point of interest of a user is realized through joint loss calculation. The application effectively improves the ranking discrimination ability of the model for similar POIs, and is suitable for personalized point of interest recommendation tasks.

Description

Technical Field

[0001] This invention belongs to the field of recommendation systems and graph neural network technology, specifically, it relates to a multi-view interest point recommendation method based on region modeling and boundary optimization. Background Technology

[0002] With the rapid development of mobile internet and location service technologies, a large amount of user trajectory data is accumulating, making next point of interest (POI) recommendation an important research direction in the field of intelligent recommendation. POI recommendation aims to predict the location a user is most likely to visit next based on their historical access sequences, thereby providing personalized location services.

[0003] Existing methods mainly include sequence-based methods, graph neural network-based methods, and hypergraph-based methods. Sequence-based methods typically use recurrent neural networks, self-attention mechanisms, or Transformer structures to model user behavior sequences, but they struggle to effectively characterize the complex high-order relationships between points of interest (POIs). Methods based on ordinary graph structures can only represent binary relationships and are unable to describe high-order co-occurrence patterns among multiple POIs. In recent years, some studies have begun to use hypergraph structures to model user behavior, connecting multiple POIs through hyperedges to enhance the representation of high-order relationships. However, existing methods still suffer from insufficient spatial region modeling capabilities, inadequate multi-view fusion, and weak negative sample optimization capabilities, making it difficult to effectively improve the performance of next POI recommendation.

[0004] Therefore, a next point of interest recommendation method that can simultaneously model regional spatial structure, multi-view semantic relationships, and boundary optimization mechanisms is needed. Summary of the Invention

[0005] To address the shortcomings of existing technologies, such as insufficient spatial region modeling, inadequate multi-view fusion, and weak sorting boundary optimization capabilities, this invention proposes a multi-view point of interest recommendation method based on region modeling and boundary optimization. This method can filter multiple points of interest that meet the personalized needs of target users, adapting to the needs of users recommending places of interest in daily life scenarios such as travel and exploring shops. This invention is achieved through the following technical solutions: A multi-view point of interest recommendation method based on region modeling and boundary optimization: The method specifically includes the following steps: Step 1: Obtain user trajectory data from the NYC and TKY public check-in datasets, sort user check-in records by access time, and construct user historical access sequences; Step 2: Construct spatial view, temporal view and collaboration view from three dimensions: space, time and user collaboration. The spatial view hypergraph models the spatial proximity relationship between points of interest, the temporal view hypergraph models the temporal transfer relationship of user access behavior, and the collaboration view hypergraph models the higher-order co-occurrence relationship between users. Step 3: Use a hypergraph convolutional network to aggregate and extract information from high-order related features of each view to obtain the interest point representation and user representation under each view; Step 4: A hierarchical progressive fusion mechanism is adopted to complete the fusion of multi-view representations, construct a cross-view contrastive learning task, and use a loss function to enhance the semantic consistency between different views. The total loss of contrastive learning is composed of the POI part loss and the user part loss. Step 5: Use a conditional diffusion model to generate boundary-hard negative samples, construct a boundary-aware ranking system to optimize the recommendation ranking effect of the next point of interest, and output the point of interest that the user is most likely to visit in the next moment after training.

[0006] Further, in step 1, The dataset includes user ID, point of interest ID, access timestamp, point of interest category, and POI geographic coordinate information.

[0007] Further, step 2 includes, Step 2.1: For each user, count the Top-K points of interest with the highest historical access frequency as central nodes to form the central node set for that user; Step 2.2: Using each central node as an anchor point, the Haversine method is used to calculate the spatial distance between the central node and the other points of interest, and the spatial distance is used to describe the geographical proximity relationship between the points of interest. Set multiple spatial distance scales and set corresponding distance thresholds for each scale in the geohash grid. When the spatial distance between the point of interest and the center node does not exceed the corresponding threshold, the point of interest is added to the spatial hyperedge of the corresponding scale. Step 2.3: Apply a Gaussian function to the nodes within the hyperedge to apply distance decay weighting, introduce multi-scale Gaussian weights, so that interest points that are closer to each other have higher spatial correlation weights, and define the weighted correlation matrix of the spatial hypergraph based on the hyperedge structure and weights.

[0008] Furthermore, in step 3, When using a hypergraph convolutional network to process different views, multi-layer propagation updates are performed based on the node degree matrix, hyperedge degree matrix, hyperedge weight matrix, and learnable parameter matrix. After multiple layers of propagation, spatial view POI representation, temporal view POI representation, and collaborative view POI representation are obtained; Under each view, the user representation under the corresponding view is obtained by aggregating the embeddings of the POIs accessed by the user.

[0009] Furthermore, step 4 includes, Step 4.1: Perform global average pooling on the POI embedding of each view to obtain a view-level semantic representation. Use cosine similarity to calculate the semantic similarity between different views and construct a similarity matrix. Iteratively select the two views with the highest similarity to merge first. After obtaining a new POI representation, update the similarity set and recalculate the similarity between the new view and the remaining views until all views are merged into a unified POI representation. Step 4.2, construct the view Figure 1 Consistency loss, constraint fusion representation maintains fidelity to the original view in the semantic space; Step 4.3: For user information under different views, a gating mechanism is introduced to learn the weights of different views, and the final user representation is obtained through weighted aggregation.

[0010] Furthermore, step 5 includes, Step 5.1: Using the embedding after multi-view fusion as the diffusion condition, Gaussian noise is gradually injected into the representation through forward diffusion in the candidate subspace to perform random perturbation; the model learns a conditional denoising network, and recovers the target representation by predicting noise to achieve the reverse denoising process; the conditional semantics guides the diffusion model to generate hard negative samples near the boundary. Step 5.2: Train the diffusion network to learn to predict noise. The diffusion model is trained by minimizing the noise prediction error. The diffusion loss is used to optimize the diffusion generation capability of the model, resulting in hard negative samples that are semantically consistent with the original POI representation but have local perturbations and are distributed near the decision boundary. Step 5.3: Construct a set of hard negative samples based on the generated samples and the candidate set, and optimize it in combination with the ranking loss; introduce geometric adaptive margin to make the negative samples that are closer to the positive samples more constrained, forming a clearer ranking boundary in the embedding space.

[0011] A multi-view point of interest recommendation system based on region modeling and boundary optimization; The system includes a sorting module, a hypergraph modeling module, a feature extraction module, a fusion module, and a recommendation module; The sorting module is used to obtain user trajectory data from the NYC and TKY public check-in datasets, sort user check-in records by access time, and construct user historical access sequences. The hypergraph modeling module constructs spatial views, temporal views, and collaborative views from three dimensions: space, time, and user collaboration. The spatial view hypergraph models the spatial proximity relationships between points of interest, the temporal view hypergraph models the temporal transition relationships of user access behavior, and the collaborative view hypergraph models the higher-order co-occurrence relationships between users. The feature extraction module uses a hypergraph convolutional network to perform information aggregation and feature extraction of high-order related features of each view, and obtains the interest point representation and user representation under each view; The fusion module adopts a hierarchical progressive fusion mechanism to complete the fusion of multi-view representations, constructs a cross-view contrastive learning task, and uses a loss function to enhance the semantic consistency between different views. The total loss of contrastive learning is composed of the POI part loss and the user part loss. The recommendation module uses a conditional diffusion model to generate boundary-hard negative samples, constructs a boundary-aware ranking system to optimize the recommendation ranking effect of the next point of interest, and outputs the point of interest that the user is most likely to visit at the next moment after training.

[0012] A computer device system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method. A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0013] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the method described above.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention can effectively solve the problems of insufficient spatial modeling, inadequate multi-view fusion, and weak sorting and discrimination capabilities in existing technologies. It can accurately mine user interests and preferences and achieve high-quality interest point recommendations, which meets the application needs of recommending multiple high-quality interest points to users in real-world scenarios.

[0015] First, this invention constructs a multi-view modeling framework, building independent hypergraphs from three dimensions: spatial, temporal, and collaborative, forming three types of preference representation networks: spatial proximity, temporal transition, and collaborative co-occurrence. It innovatively adopts a center-driven multi-scale region coding strategy, using frequently accessed Points of Interest (POIs) as core anchors to capture regional spatial dependency patterns at different scales, significantly enhancing spatial feature representation capabilities.

[0016] Secondly, this invention innovatively proposes a hierarchical alignment mechanism that integrates progressive feature fusion with a Huffman tree-like greedy strategy and cross-view comparative learning. It first integrates spatial, temporal, and collaborative view features based on semantic similarity, and then achieves semantic alignment of multi-view embeddings through comparative constraints. This efficiently integrates complementary information from each view, resulting in a unified, accurate, and robust user representation, avoiding semantic interference and information loss problems caused by direct multi-view fusion.

[0017] Finally, this invention proposes a dynamic ranking boundary enhancement strategy, which generates boundary-aware hard-negative samples in the candidate subspace and optimizes them by combining dynamic margin ranking loss, effectively improving the model's ability to rank and distinguish semantically similar POIs.

[0018] This invention is applicable to personalized point-of-interest recommendation tasks; it can be applied to scenarios such as location recommendation, smart cities, mobile services, and personalized recommendations. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Unless otherwise specified, the experimental methods used in the following examples are conventional methods. Unless otherwise specified, the materials, reagents, methods, and instruments used are all conventional materials, reagents, methods, and instruments in the art, and can be obtained commercially by those skilled in the art.

[0022] like Figure 1 As shown, a multi-view interest point recommendation method based on region modeling and boundary optimization includes the following steps: Step 1: Obtain user trajectory data from the NYC and TKY public check-in datasets. These datasets include user IDs, Point of Interest (POI) IDs, access timestamps, POI (Next Point-of-Interest) categories, and POI geographic coordinates. Sort user check-in records according to access time and construct a user historical access sequence.

[0023] The user set in the training set is , ,in The total number of users; the set of points of interest is , ,in This represents the total number of POIs.

[0024] For any user Its historical access sequence Defined as ,in Indicates the user at a certain time. POIs visited Indicates the length of the user access sequence.

[0025] The next point of interest recommendation task is defined as predicting the point of interest that a user is most likely to visit at the next moment based on the user's historical access sequence.

[0026] Step 2: To comprehensively explore the association features between POIs and users from three different dimensions—space, time, and user collaboration—and to finely characterize multi-source dependencies, a spatial view is used to model the spatial proximity relationships between POIs, a temporal view is used to model the temporal transition relationships of user access behavior, and a collaboration view is used to model the higher-order co-occurrence relationships between users. A hypergraph for the spatial view is constructed accordingly. Time view hypergraph and collaborative view hypergraph Define the hypergraph structure as follows: ,in, Represents a set of nodes. Denotes the set of superedges. This represents the hypergraph incidence matrix.

[0027] Step 2-1, for each user The top-K points of interest with the highest historical visit frequency of users are selected as the central nodes. .in, Indicates user The set of central nodes, Indicates the frequency of visits to points of interest. Indicates the number of central nodes. Indicates the first A high-frequency POI.

[0028] Step 2-2: These centers reflect the user's core regional interests. (Using each anchor point...) Centered on, Hasersine is used to compute the central node and other points of interest. Spatial distance between , .

[0029] in, Indicates latitude, Indicates longitude. This represents the Earth's radius.

[0030] Spatial distance is used to describe the geographic proximity between points of interest. Multiple spatial distance scales are set, and multi-scale distance thresholds are configured within the geohashing grid. , where S is the partitioning scale. When satisfying At that time, the point of interest is added to the corresponding scale space hyperedge. Enhance the ability to model regional spatial relationships by using multi-scale spatial hyperedges.

[0031] Steps 2-3: To characterize the effect of distance decay, Gaussian functions are applied to the nodes inside the hyperedge to weight the distance decay, introducing multi-scale Gaussian weights. , .

[0032] in, This represents the Gaussian kernel width parameter. Interest points that are closer together have higher spatial association weights. Finally, based on the above hyperedge structure and weights, a weighted association matrix for the spatial hypergraph is defined.

[0033] Step 3: In order to achieve information aggregation and feature extraction of higher-order related features, and to fully explore higher-order dependencies such as spatial adjacency, temporal transition, and user collaboration, a hypergraph convolutional network is used to process different views.

[0034] The hypergraph convolution update formula is defined as follows: .in, For the first Layer node representation, Represents the node degree matrix, Represents the hypermarginality matrix. Represents the hyperedge weight matrix. This represents the learnable parameter matrix.

[0035] After multiple layers of propagation, the spatial view POI representation is obtained. Time view POI representation and collaborative view POI representation .

[0036] For user feedback Aggregation is performed under each view by accessing the POI's embedding. ,in .

[0037] Step 4: After obtaining the embedded representation, a hierarchical progressive fusion mechanism is used to fuse multi-view representations and construct a cross-view comparison learning task.

[0038] Cosine similarity is used to calculate the semantic similarity between different views. Formulated as .in, and These represent the node representations under different views.

[0039] By dynamically selecting the views with the highest semantic similarity and progressively fusing them, a unified fused representation is obtained.

[0040] Using InfoNCE loss to enhance semantic consistency between different views .in, Indicates a positive sample. Represents the set of negative samples. This represents the temperature parameter.

[0041] Contrastive learning enhances the consistency of representations across different views. Ultimately, the total loss from contrastive learning is... .

[0042] Step 4-1, for each view First, global average pooling is performed on the POI embeddings to obtain the view-level semantic representation. , ,in Indicates the first The first view Embedding vectors of POIs.

[0043] Cosine similarity is used to calculate the semantic similarity between different views. And construct a similarity matrix , .

[0044] The two most similar views are selected for priority merging. After obtaining the new POI representation through the fusion function, the similarity set is updated. , .in, The fusion function (such as MLP) is used, and the similarity between the new view and the remaining views is recalculated step by step. This process is repeated until all views are merged into a unified representation. This reduces semantic conflicts caused by direct merging of multiple views through a progressive approach.

[0045] Step 4-2: To avoid the destruction of information in any view during the merging process, the merged representation maintains consistency with the original view. Therefore, construct the view Figure 1 Sexual damage This constraint encourages fusion representations to maintain fidelity to the original view in the semantic space, thereby improving the stability and discriminative power of the representation.

[0046] Step 4-3: In order to integrate user information from different views, a gating mechanism is introduced to learn the weights of different views. , To obtain end-user feedback , .

[0047] Step 5: Generate boundary hard-to-bear samples using a conditional diffusion model and construct a boundary-aware ranking optimization objective.

[0048] The forward diffusion process is defined as... .in, Indicates the first Step diffusion state, This represents the diffusion noise figure.

[0049] Difficult negative samples that are semantically close to the real target interest points are generated through a reverse denoising process. The overall optimization objective is defined as follows: .in, Indicates recommended loss. Indicates fusion loss, Indicates the contrast learning loss. Indicates diffusion loss, This represents the loss due to poor sample ranking. The trained model outputs the points of interest that the user is most likely to visit in the next time step.

[0050] Step 5-1, embedding the multi-view merged data. As a diffusion condition, to ensure the semantic consistency of the perturbation, the diffusion process first injects Gaussian noise into the representation through forward diffusion in the candidate subspace to randomly perturb it.

[0051] Subsequently, the model learns a conditional denoising network, which recovers the target representation through the reverse denoising process by predicting noise. ,in: This represents conditional semantics. Conditional semantics guides the diffusion model to generate hard-negative samples near the boundary.

[0052] Step 5-2: Train the diffusion network to learn how to predict noise. The diffusion model is trained by minimizing the noise prediction error and optimized using diffusion loss. .in, Represents real noise. This represents the predicted noise.

[0053] The diffusion generation capability is optimized by leveraging noise prediction errors. Through this process, the model can generate new samples that are semantically consistent with the original POI representation but have local perturbations. These samples are typically distributed near the decision boundary, thus constituting potential hard-to-bear samples. .

[0054] Step 5-3: Based on the generated samples and the candidate set, construct the hard-to-bear sample set. And optimize by combining the ranking loss.

[0055] Introducing geometrically adaptive margins, ,in For the first Embedding vectors of candidate negative sample POIs This is the embedding vector of the positive sample POI. The final optimization goal is to , This mechanism imposes stronger constraints on negative samples that are closer to positive samples, thereby creating clearer ranking boundaries in the embedding space.

[0056] A computer device system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method. A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0057] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the method described above.

[0058] The memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory of the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0059] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means such as coaxial cable, optical fiber, digital subscriber line, DSL, or wireless means such as infrared, wireless, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium such as a floppy disk, hard disk, magnetic tape; an optical medium such as a high-density digital video disc, DVD; or a semiconductor medium such as a solid-state disk, SSD, etc.

[0060] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0061] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as execution by a hardware decoding processor, or as execution by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0062] The above provides a detailed description of the multi-view interest point recommendation method based on region modeling and boundary optimization proposed in this invention, and elucidates the principles and implementation methods of this invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A multi-view point of interest recommendation method based on region modeling and boundary optimization, characterized in that: The method specifically includes the following steps: Step 1: Obtain user trajectory data from the NYC and TKY public check-in datasets, sort user check-in records by access time, and construct user historical access sequences; Step 2: Construct spatial view, temporal view and collaboration view from three dimensions: space, time and user collaboration. The spatial view hypergraph models the spatial proximity relationship between points of interest, the temporal view hypergraph models the temporal transfer relationship of user access behavior, and the collaboration view hypergraph models the higher-order co-occurrence relationship between users. Step 3: Use a hypergraph convolutional network to aggregate and extract information from high-order related features of each view to obtain the interest point representation and user representation under each view; Step 4: A hierarchical progressive fusion mechanism is adopted to complete the fusion of multi-view representations, construct a cross-view contrastive learning task, and use a loss function to enhance the semantic consistency between different views. The total loss of contrastive learning is composed of the POI part loss and the user part loss. Step 5: Use a conditional diffusion model to generate boundary-hard negative samples, construct a boundary-aware ranking system to optimize the recommendation ranking effect of the next point of interest, and output the point of interest that the user is most likely to visit in the next moment after training.

2. The recommended method according to claim 1, characterized in that: In step 1, The dataset includes user ID, point of interest ID, access timestamp, point of interest category, and POI geographic coordinate information.

3. The recommended method according to claim 2, characterized in that: Step 2 includes, Step 2.1: For each user, count the Top-K points of interest with the highest historical access frequency as central nodes to form the central node set for that user; Step 2.2: Using each central node as an anchor point, the Haversine method is used to calculate the spatial distance between the central node and the other points of interest, and the spatial distance is used to describe the geographical proximity relationship between the points of interest. Set multiple spatial distance scales and set corresponding distance thresholds for each scale in the geohash grid. When the spatial distance between the point of interest and the center node does not exceed the corresponding threshold, the point of interest is added to the spatial hyperedge of the corresponding scale. Step 2.3: Apply a Gaussian function to the nodes within the hyperedge to apply distance decay weighting, introduce multi-scale Gaussian weights, so that interest points that are closer to each other have higher spatial correlation weights, and define the weighted correlation matrix of the spatial hypergraph based on the hyperedge structure and weights.

4. The recommended method according to claim 3, characterized in that: In step 3, When using a hypergraph convolutional network to process different views, multi-layer propagation updates are performed based on the node degree matrix, hyperedge degree matrix, hyperedge weight matrix, and learnable parameter matrix. After multiple layers of propagation, spatial view POI representations, temporal view POI representations, and collaborative view POI representations are obtained; Under each view, the user representation under the corresponding view is obtained by aggregating the embeddings of the POIs accessed by the user.

5. The recommended method according to claim 4, characterized in that: Step 4 includes, Step 4.1: Perform global average pooling on the POI embedding of each view to obtain the view-level semantic representation. Use cosine similarity to calculate the semantic similarity between different views and construct a similarity matrix. Iteratively select the two views with the highest similarity to merge first, and after obtaining a new POI representation, update the similarity set, recalculate the similarity between the new view and the remaining views, until all views are merged into a unified POI representation; Step 4.2: Construct the view consistency loss and constrained fusion representation to maintain fidelity to the original view in the semantic space; Step 4.3: For user information under different views, a gating mechanism is introduced to learn the weights of different views, and the final user representation is obtained through weighted aggregation.

6. The recommended method according to claim 5, characterized in that: Step 5 includes, Step 5.1: Using the embedding after multi-view fusion as the diffusion condition, Gaussian noise is gradually injected into the representation through forward diffusion in the candidate subspace to perform random perturbation; The model learns a conditional denoising network, which recovers the target representation by predicting noise and uses the conditional semantics to guide the diffusion model to generate hard negative samples near the boundary. Step 5.2: Train the diffusion network to learn to predict noise. The diffusion model is trained by minimizing the noise prediction error. The diffusion loss is used to optimize the diffusion generation capability of the model, resulting in hard negative samples that are semantically consistent with the original POI representation but have local perturbations and are distributed near the decision boundary. Step 5.3: Construct a set of hard negative samples based on the generated samples and the candidate set, and optimize it in combination with the ranking loss; introduce geometric adaptive margin to make the negative samples that are closer to the positive samples more constrained, forming a clearer ranking boundary in the embedding space.

7. A multi-view point of interest recommendation system based on region modeling and boundary optimization, characterized in that: The system is used to execute the multi-view point of interest recommendation method based on region modeling and boundary optimization as described in any one of claims 1 to 6; The system includes a sorting module, a hypergraph modeling module, a feature extraction module, a fusion module, and a recommendation module; The sorting module is used to obtain user trajectory data from the NYC and TKY public check-in datasets, sort user check-in records by access time, and construct user historical access sequences. The hypergraph modeling module constructs spatial views, temporal views, and collaborative views from three dimensions: space, time, and user collaboration. The spatial view hypergraph models the spatial proximity relationships between points of interest, the temporal view hypergraph models the temporal transition relationships of user access behavior, and the collaborative view hypergraph models the higher-order co-occurrence relationships between users. The feature extraction module uses a hypergraph convolutional network to perform information aggregation and feature extraction of high-order related features of each view, and obtains the interest point representation and user representation under each view; The fusion module adopts a hierarchical progressive fusion mechanism to complete the fusion of multi-view representations, constructs a cross-view contrastive learning task, and uses a loss function to enhance the semantic consistency between different views. The total loss of contrastive learning is composed of the POI part loss and the user part loss. The recommendation module uses a conditional diffusion model to generate boundary-hard negative samples, constructs a boundary-aware ranking system to optimize the recommendation ranking effect of the next point of interest, and outputs the point of interest that the user is most likely to visit at the next moment after training.

8. A computer device system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program / instructions, characterized in that, When executed by a processor, the computer program instructions implement the steps of the method according to any one of claims 1 to 6.

Images