Method for recommending next interest point based on cross-regional city space knowledge graph

Abstract

The invention belongs to the technical field of knowledge maps, and discloses a next interest point recommendation method based on a cross-regional city space knowledge map. According to the method, on the basis of the geographic space knowledge graph and the user preference knowledge graph, a cross-regional relationship modeling mechanism is introduced, the spatial proximity and the reachability between the regions are comprehensively considered, and more comprehensive description of the cross-regional behavior of the user is realized. According to the method, a recommendation framework combining a geographic module and a sequence module is designed, spatial dependence between interest points can be captured, a high-order sequence mode in user sign-in behaviors can be mined, and recommendation accuracy and diversity are improved. Geographic representation and sequence representation are fused through a consistency learning framework, the robustness and generalization ability of the model can be enhanced, and the model can still keep stable performance in sparse data and cold start scenes. According to the method, the defects of insufficient region boundary perception, recommendation result centralization, poor cross-region prediction adaptability and the like of an existing method are overcome.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph technology, specifically relating to a method for recommending next-point-of-interest (POI) based on cross-regional urban spatial knowledge graphs, which is particularly suitable for recommending next-point-of-interest (POI) in cross-regional scenarios. Background Technology

[0002] Existing methods for constructing urban spatial knowledge graphs mainly fall into two categories: point-based modeling methods and region-based modeling methods. The former uses the latitude and longitude coordinates of Points of Interest (POIs) as the core, constructing the graph structure by calculating location proximity relationships; the latter uses administrative boundaries or functional zones as the basis, dividing POIs into different regions to express the spatial structure in a hierarchical manner. Furthermore, researchers have proposed a method that integrates multi-source heterogeneous information, incorporating attributes such as POI functional types and user check-in data into the graph modeling, thereby forming a composite knowledge graph that combines spatial structure and behavioral characteristics. However, existing urban spatial knowledge graph construction methods have weak adaptability in cross-regional scenarios. Due to the lack of effective modeling of regional boundary information, spatial accessibility, and the co-occurrence relationships of user cross-regional behaviors, the graph's expressive power in cross-regional travel scenarios is insufficient.

[0003] Existing Next-Point-of-Interest (Next-POI) recommendation methods can be broadly categorized into three types: traditional methods, sequence modeling methods, and graph neural network methods. Traditional methods, including collaborative filtering and Markov chain models, utilize users' historical browsing trajectories for prediction. Sequence modeling methods, such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), can capture the temporal dependencies of user behavior sequences. Graph neural network-based methods incorporate spatial relationships and semantic attributes between points of interest into the model, leveraging graph structures to enhance recommendation performance. However, existing Next-POI recommendation methods generally suffer from a "centralization" problem, with recommendations overly concentrated in popular areas, resulting in insufficient coverage of peripheral areas and low-frequency points of interest. Furthermore, in scenarios with sparse data and cold start conditions, the models do not adequately characterize users' true travel intentions.

[0004] Furthermore, existing methods lack adaptability in cross-regional recommendation scenarios, primarily due to their difficulty in characterizing behavioral transfer patterns and cross-regional access patterns, leading to decreased prediction accuracy and poor generalization ability. Therefore, there is an urgent need to propose a graph construction method that can integrate inter-regional relationships and model cross-regional behavior to improve the graph's completeness and applicability. In addition, a recommendation method that can alleviate the regional centralization problem and maintain robustness in sparse data scenarios is needed. Simultaneously, a method that can comprehensively consider geospatial structure and user behavior sequence features within a unified framework is also required to improve the accuracy and stability of cross-regional recommendations. Summary of the Invention

[0005] The purpose of this invention is to propose a next point of interest recommendation method based on a cross-regional urban spatial knowledge graph. This method proposes a modeling and recommendation approach that integrates cross-regional relationships for urban spatial knowledge graph and next point of interest recommendation scenarios, so as to realize functions such as cross-regional knowledge graph construction, user behavior modeling and next point of interest prediction. This helps to alleviate the shortcomings of existing methods such as insufficient perception of regional boundaries, centralized recommendation results and poor cross-regional prediction adaptability.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The next point of interest recommendation method based on cross-regional urban spatial knowledge graphs includes the following steps: Step 1. Construct a scalable geospatial knowledge graph based on massive urban spatial data; perform spatial data cleaning by combining administrative division data and user check-in data; and generate a spatial relationship structure between points of interest based on this data. The spatial relationship structure includes spatial edges of spatial proximity relationships and cross-regional relationships. Spatial proximity relationships are modeled using a unified fixed weight, while cross-regional spatial relationships are modeled using an inter-regional heat matrix. Differentiated weights are assigned to cross-regional edges, and dynamic updates of the weights are supported to characterize the differences in migration intensity between regions. Step 2. Construct a sequence graph based on the time sequence of user check-in data to generate a user preference knowledge graph that characterizes user behavior preferences and migration patterns, and combine it with the constructed geospatial knowledge graph to carry out cross-regional knowledge graph fusion; Step 3. Build a next point of interest recommendation model that includes a geographic module, a sequence module, and a preference fusion prediction module; the geographic module and the sequence module are both implemented based on a graph neural network architecture; The geography module uses graph neural networks to model spatial preferences for spatial proximity relationships and cross-regional accessibility in the fused knowledge graph, in order to extract users' spatial behavior characteristics at different spatial scales; The sequence module performs sequence preference modeling based on the user access sequence depicted in the fused knowledge graph. It learns long-term user preferences and high-order travel patterns through attention mechanisms and random walk structures to perform preference fusion prediction. The preference fusion prediction module jointly models spatial preference representation and sequential preference representation through a fusion mechanism to generate the probability distribution of user access to potential points of interest, thereby enabling the recommendation of the next point of interest in the user's mobile behavior prediction task.

[0007] Furthermore, based on the aforementioned method for recommending next point of interest (OPI) based on cross-regional urban spatial knowledge graphs, this invention also proposes a corresponding OPI recommendation system based on cross-regional urban spatial knowledge graphs, the scheme of which is as follows: The next point of interest recommendation system based on cross-regional urban spatial knowledge graph includes the following modules: The module for constructing geospatial knowledge graphs and modeling cross-regional dynamic relationships is used to build scalable geospatial knowledge graphs for massive urban spatial data; it performs spatial data cleaning by combining administrative division data and user check-in data; on this basis, it generates spatial relationship structures between points of interest and constructs spatial edges that include spatial proximity relationships and cross-regional relationships. A unified fixed weight is used to model spatial proximity relationships, while an inter-regional heat matrix is ​​introduced for cross-regional spatial relationships. Differentiated weights are assigned to cross-regional edges, and dynamic updates of the weights are supported to characterize the differences in migration intensity between regions. The user preference knowledge graph construction and cross-regional knowledge graph fusion module is used to construct a sequence graph based on the time sequence of user check-in data, generate a user preference knowledge graph that describes user behavior preferences and migration patterns, and carry out cross-regional knowledge graph fusion in combination with the constructed geospatial knowledge graph. And a next point of interest recommendation module, used to build a next point of interest recommendation model that includes a geographic module, a sequence module, and a preference fusion prediction module; the geographic module and the sequence module are both implemented based on a graph neural network architecture; The geography module uses graph neural networks to model spatial preferences for spatial proximity and cross-regional relationships in the fused knowledge graph, in order to extract users' spatial behavior characteristics at different spatial scales. The sequence module performs sequence preference modeling based on the user access sequence depicted in the fused knowledge graph. It learns long-term user preferences and high-order travel patterns through attention mechanisms and random walk structures to perform preference fusion prediction. The preference fusion prediction module jointly models spatial preference representation and sequential preference representation through a fusion mechanism to generate the probability distribution of user access to potential points of interest, thereby enabling the recommendation of the next point of interest in the user's mobile behavior prediction task.

[0008] The present invention has the following advantages: As described above, this invention discloses a next-point-of-interest (POI) recommendation method based on a cross-regional urban spatial knowledge graph. In terms of urban spatial modeling, this invention can automatically construct dynamic relationships between regions from continuous user check-in data and achieve adaptive updates of cross-regional edge weights based on a heat matrix, thereby forming a more realistic spatial interaction structure and providing effective support for regional boundary and cross-regional travel prediction. Regarding graph-level fusion, this invention unifies the user preference knowledge graph and the geospatial knowledge graph into a single structure through node alignment and edge-level fusion, forming a multi-relationship urban knowledge graph that simultaneously includes proximity relationships, cross-regional relationships, and access sequence relationships, achieving a structurally consistent and semantically complete urban knowledge representation. In terms of recommendation model structure, this invention employs graph neural networks, random walk graph kernels, and multi-head attention mechanisms to achieve deep fusion of spatial preferences and behavioral preferences, capturing users' high-order behavioral patterns and spatial dependency structures, providing end-to-end high-performance prediction capabilities for Next-POI recommendations. In terms of data, this invention processes urban spatial data, administrative division data, and user mobility behavior data, enabling unified processing of multi-source heterogeneous urban data simultaneously. It possesses high application value for smart mobility and urban service scenarios. Functionally, this invention achieves high-precision Next-POI recommendations, combining spatial structure, cross-regional relationships, and behavioral sequences for joint modeling. It exhibits significant predictive advantages on real-world urban data and has practical deployment value. Regarding robustness, this invention maintains good recommendation performance even under adverse conditions such as data sparsity and user cold start. Through cross-regional knowledge fusion and consistency learning, it enhances model generalization, making it suitable for various practical application scenarios. Attached Figure Description

[0009] Figure 1 This is a technical roadmap for the next point of interest recommendation method based on cross-regional urban spatial knowledge graphs of the present invention. Figure 2 This diagram illustrates the specific technical steps of Next-POI user mobility behavior prediction in this invention. Figure 3 This is a schematic diagram comparing the performance of various models under different training data volumes in a specific example of the present invention; Figure 4 This is a schematic diagram comparing the performance of various models in the ablation experiment in a specific example of the present invention. Detailed Implementation

[0010] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 This embodiment 1 describes a next point of interest recommendation method based on cross-regional urban spatial knowledge graphs to alleviate the shortcomings of existing methods in terms of insufficient regional boundary perception, centralized recommendation results, and poor cross-regional prediction adaptability.

[0011] The present invention is mainly improved in the following three aspects, specifically: 1. In terms of cross-regional knowledge graph construction: Based on geospatial knowledge graphs and user preference knowledge graphs, this invention introduces a cross-regional relationship modeling mechanism, comprehensively considering spatial proximity and inter-regional accessibility to achieve a more comprehensive description of users' cross-regional behavior.

[0012] 2. Next point of interest recommendation method: This invention designs a recommendation framework that combines a geographic module and a sequence module, which can capture the spatial dependencies between points of interest and mine high-order sequence patterns in user check-in behavior, thereby improving the accuracy and diversity of recommendations.

[0013] 3. Enhanced adaptability of cross-regional recommendations: This invention integrates geographic representation and sequence representation through a consistency learning framework, which enhances the robustness and generalization ability of the model, enabling it to maintain stable performance even under sparse data and cold start scenarios.

[0014] like Figure 1 As shown, the next point of interest recommendation method based on a cross-regional urban spatial knowledge graph includes the following steps: Step 1. Construction of geospatial knowledge graph and modeling of cross-regional dynamic relationships.

[0015] To address massive urban spatial data, we construct an scalable geospatial knowledge graph; we perform spatial data cleaning by combining administrative division data and user check-in data; and on this basis, we generate a spatial relationship structure between points of interest.

[0016] The spatial relationship structure includes spatial boundaries that encompass both spatial proximity relationships and cross-regional relationships.

[0017] A unified fixed weight is used to model spatial proximity relationships, while an inter-regional heat matrix is ​​introduced for cross-regional spatial relationships. Differentiated weights are assigned to cross-regional edges, and dynamic updates of the weights are supported to characterize the differences in migration intensity between regions.

[0018] Step 1.1. Administrative region screening and spatial object cleaning.

[0019] Step 1.1 involves spatial data cleaning by combining administrative division data with user check-in data. Only administrative regions with user check-in activity are retained, and invalid or redundant spatial objects are removed. The process is as follows: Construct a region set based on administrative boundary data. By combining user check-in data, it can be determined whether each administrative region contains valid check-in points. Only administrative regions containing at least one check-in point are retained, and all other regions are considered invalid and removed, thereby avoiding redundant regions from causing noise to the map structure.

[0020] Filtered set of administrative regions .

[0021] in Represents a POI in geospatial space. This represents the set of POIs in the user check-in dataset. Indicates the use of POI for determining The region to which it belongs. This refers to the original object in the OpenStreetMap data source. Only spatial objects located within the effective administrative area are retained. This process removes objects with missing coordinates, missing attributes, or no names, ensuring that all final POIs have stable identification and geographic availability.

[0022] Generate a uniform format for the preserved spatial objects, including the name field. Centroid latitude and longitude coordinates Functional attribute tags Used for subsequent spatial relationship construction; where the centroid coordinates Calculated using the following formula: ; in Number of boundary points: if it is a point object =1; if it is a road object or a region object =Number of outline points of the object; , These represent the first and second parts of the object. The latitude and longitude of each boundary point.

[0023] The cleaned POI data is retained and denoted as a geospatial knowledge graph. POI set .

[0024] Step 1.2. Generation of spatial proximity and cross-regional relationships.

[0025] For any two POIs, define them respectively as follows: , The spatial relationship between the two can be determined in the following way: ; in express and Spatial relationships between them; This represents a spatial proximity edge generated within the same region. This indicates a cross-regional relationship edge.

[0026] pass The function checks if objects are in the same domain. and Within the same domain and at a distance less than or equal to the nearest neighbor threshold If they are spatially adjacent, then they are spatially adjacent; if and Foreign regions and distance less than or equal to the cross-region threshold This indicates a cross-regional relationship.

[0027] All eligible spatial proximity relationships and cross-regional relationships will be added to the spatial relationship set. This leads to the formation of a geospatial knowledge graph. This provides a basis for subsequent fusion maps and GNN inputs.

[0028] Step 1.3. Spatial relationship weight setting.

[0029] Spatial proximity edges generated within the same region Since it is only used to represent local spatial reachability, all neighboring edges are assigned a fixed weight of 1. This fixed weight is used to ensure that the proximity relationships have a consistent structural representation in the graph, avoiding bias caused by uneven data distribution.

[0030] For cross-regional relationship edges The weights of regions are dynamically assigned based on their popularity; and cross-regional access is statistically analyzed based on users' consecutive check-in sequences to construct a regional popularity matrix. : ; in For the number of regions, This indicates that the user's consecutive check-ins are from the region. To the region The number of cross-regional behaviors.

[0031] Cross-regional relationship weight Defined as: ; in For the region The maximum relation weight is the global scaling factor; Indicates the area The target region with the highest popularity among all cross-regional activities; weighted by scaling. Limited to Inside.

[0032] Specifically, the maximum relation weight for: ; in Indicates from the region Number of cross-regional activities to other regions.

[0033] Indicates all regions, from region The maximum number of cross-regional trips to other regions; this ratio reflects the region's... The value is taken from the perspective of the overall impact on travel. Within the range.

[0034] Through the weight setting process, structurally consistent, behavior-driven, and dynamically adjustable edge weights are generated for all spatial relationships, thereby providing high-quality spatial input for subsequent cross-regional knowledge graph construction and graph neural network modeling.

[0035] This invention addresses massive urban spatial data by constructing a scalable geospatial knowledge graph model. It combines administrative division data with user check-in data for spatial data cleaning, retaining only administrative regions with user check-in activity and removing invalid or redundant spatial objects. Based on this, a spatial relationship structure between Points of Interest (POIs) is generated, constructing spatial edges that include both spatial proximity and cross-regional spatial relationships. Spatial proximity relationships are modeled using uniform, fixed weights, while cross-regional spatial relationships are modeled using an inter-regional heat matrix. Differentiated weights are assigned to cross-regional edges, and dynamic updates of these weights are supported to characterize differences in migration intensity between regions. This provides a precise and updatable spatial structure foundation for cross-regional knowledge fusion and user behavior modeling.

[0036] Step 2. Construction of user preference knowledge graph and fusion of cross-regional knowledge graph.

[0037] This invention constructs a sequence graph based on the time sequence of user check-in data, generates a user preference knowledge graph that characterizes user behavior preferences and migration patterns, and combines it with the constructed geospatial knowledge graph to carry out cross-regional knowledge graph fusion. The user preference knowledge graph and the geospatial knowledge graph are merged at the node and edge levels to form a unified cross-regional knowledge graph.

[0038] Step 2.1. Construction of the user preference knowledge graph.

[0039] For users Sign-in data Perform parsing and extract the user's information. Collection of POIs accessed And construct directed action edges for adjacent POIs that check in. .

[0040] in express The number of times you sign in. Indicates user No. The POI for the first check-in, and the two consecutive check-ins , There is a user preference relationship. Merge all relationships to obtain set of preference relations .

[0041] Thus, this user was formed. User Preference Knowledge Graph .

[0042] After integrating the preference paths of all users, a complete user preference knowledge graph is constructed. .

[0043] The set of nodes For all users The check-in POIs and their sequential relationships, since different users may visit the same POIs, make the graph overlap and temporal, providing rich behavioral features for subsequent modeling.

[0044] edge set For all users This sequence graph represents the access order relationship between adjacent Points of Interest (POIs) in the check-in sequence. It is used to express the sequential dependencies of user behavior, providing a behavioral information foundation for cross-regional graph fusion.

[0045] Step 2.2. Cross-regional knowledge graph fusion.

[0046] Cross-regional knowledge graph fusion aligns and merges nodes and edges, automatically removing redundant entities and invalid edges while preserving key spatial and behavioral relationships, thus forming a unified cross-regional knowledge graph representation.

[0047] The fused knowledge graph simultaneously includes spatial proximity relationships, cross-regional relationships, and user behavior sequence relationships, constructing a structurally consistent and semantically complete multi-relational urban knowledge representation.

[0048] Specifically, the cross-regional knowledge graph fusion process in this embodiment is as follows: To unify the modeling of spatial structure and user behavior, a geospatial knowledge graph is used as a benchmark to map nodes in the user preference knowledge graph. For any POI, it is defined as... Node set Represent all user check-in POIs and their order; find the nearest node in the geospatial knowledge graph. , .

[0049] in Represents the geographical distance between any two POIs; through Choose the node with the smallest distance as the mapped node; Representing geospatial knowledge graphs The set of POIs.

[0050] After processing the node mapping, the user preference knowledge graph edge set is... Each edge in Mapped to Add it to the new edge set In, retain the appearance The nodes in the graph are used to obtain the node set of the merged knowledge graph: .

[0051] Simultaneously, node sets are selected from the geospatial knowledge graph. The spatial relationships between them yield the preserved set of geographical edges. This process removes invalid nodes and edges.

[0052] This is a set of spatial relationships consisting of all eligible spatial proximity and cross-regional relationships. The relationship between the geospatial knowledge graph and the user preference knowledge graph is merged to obtain the fused knowledge graph. It combines geospatial structure and behavioral sequence information, providing a unified input structure for GNN recommendation models.

[0053] This invention constructs a sequence graph based on the temporal order of user check-in data, generating a user preference knowledge graph that characterizes user behavior preferences and migration patterns. Building upon this, it integrates the geospatial knowledge graph constructed in step 1 to perform cross-regional knowledge graph fusion, aligning and merging nodes and edges. While preserving key spatial and behavioral relationships, it automatically removes redundant entities and invalid edges, thus forming a unified cross-regional knowledge graph representation. The fused knowledge graph simultaneously includes spatial proximity relationships, cross-regional relationships, and user behavior sequence relationships, constructing a structurally consistent and semantically complete multi-relational city knowledge representation, providing a unified input carrier for subsequent graph neural network models.

[0054] Step 3. Next-Point of Interest (POI) recommendation: Predicting user movement behavior.

[0055] like Figure 2As shown, this embodiment constructs a Next Point of Interest (POI) recommendation model comprising a geographic module, a sequence module, and a preference fusion prediction module. Based on the fused cross-regional knowledge graph, it jointly models the user's spatial behavior features and sequence preference features to achieve Next-POI recommendation, i.e., Next-POI prediction. Both the geographic and sequence modules are implemented based on a graph neural network architecture. The geographic module models spatial preferences, extracting user spatial behavior features based on spatial adjacency and cross-regional relationships in the fused graph. The sequence module models sequence preferences, constructing behavioral semantic representations based on user access sequences and extracting higher-order travel pattern representations of users. The preference fusion prediction module combines spatial and sequence preference representations, generating the final POI access probability through a fusion mechanism, thereby achieving Next-POI recommendation.

[0056] Specifically, a next point of interest recommendation model is built, which includes a geographic module, a sequence module, and a preference fusion prediction module.

[0057] The geography module uses graph neural networks to model spatial preferences for spatial proximity relationships and cross-regional accessibility in the fused knowledge graph, in order to extract users' spatial behavior characteristics at different spatial scales.

[0058] The sequence module performs sequence preference modeling based on the user access sequences depicted in the fused knowledge graph. It learns long-term user preferences and high-order travel patterns through attention mechanisms and random walk structures to perform preference fusion prediction.

[0059] The preference fusion prediction module jointly models spatial preference representation and sequential preference representation through a fusion mechanism to generate the probability distribution of user access to potential points of interest, thereby enabling the recommendation of the next point of interest in the user's mobile behavior prediction task.

[0060] The processing flow of the geographic module is as follows: In a geospatial graph neural network, each POI node The layer Represented as: .

[0061] in, For this node First-order adjacency set A node in; For nodes to its adjacent nodes The vector for message passing; For the first Layer trainable weight matrix.

[0062] ; in Indicates the weight of cross-regional relationships. This represents the distance between two nodes.

[0063] go through Layered message passing obtains all POI geographic representations. : ; For users Sign-in sequence Extract the visited POIs to obtain a geographic representation. .

[0064] Utilizing multi-head attention mechanism for users Behavioral preferences are aggregated to obtain their semantic representation. : ; in, , ; , , The first Attention header query, key-value mapping matrix For the original embedding dimension, For the number of attention heads, For POI In the The attention score in each self-attention head, i.e., the self-attention weight.

[0065] Depend on Perform queries and key-value mappings, then dot products, and finally scale the results; Softmax normalization is performed to obtain .

[0066] In Next-POI recommendations, user check-in sequences not only reflect local adjacency relationships but also implicitly contain higher-order travel patterns, such as cyclical structures like "home—subway station—office." To capture these complex patterns, a graph kernel neural network (GKNN) based on random walks is introduced into the sequence module to model the latent substructure features in the sequence graph.

[0067] The processing procedure of the sequence module is as follows: Given user User Preference Knowledge Graph .

[0068] The set of nodes The set of edges represents the points of interest (POIs) visited by the user. This indicates the directed transfer relationship between adjacent sign-ins.

[0069] For any two graphs and Its direct product graph is defined as .

[0070] in , , , The diagrams are shown below. , The set of nodes and edges of a , the set of nodes in a , or the set of edges in a , or ... edge set .

[0071] exist A random walk on is defined as simultaneously on and Synchronous walk on the surface. Let... for adjacency matrix If it is a vector of all 1s, then Step random walk kernel Used to measure the length of two graphs. Similarity along the path.

[0072] in In the product graph The upper length is The counting matrix of all random walks.

[0073] There is a set Image filter And for each user User Preference Knowledge Graph Kernel similarity calculation is performed with the graph filter; where, This represents a set of trainable graph filters.

[0074] Each graph filter corresponds to a parameterized small-scale graph structure used to match high-order sequence patterns in the user sequence graph.

[0075] For the set of random walk step sizes The characteristic matrix is ​​obtained. .

[0076] The Middle The row corresponds to the first The similarity vector of the graph filter, the first Column corresponding random walk step size .

[0077] Feature matrix matrix elements Defined as: ; matrix Flatten and stitch together before inputting into a fully connected layer. Get users Sequence semantic representation ; This indicates a splicing operation. It represents a higher-order behavioral preference vector of users in the sequence dimension, which can simultaneously reflect local transition relationships and global structural patterns, providing semantic representation support for subsequent recommendation tasks.

[0078] The preference fusion prediction module uses a multilayer perceptron (MLP) to jointly model the user's geospatial preference representation and sequential semantic representation to calculate the predicted probability of the user visiting a point of interest (POI). Since the geographic module and the sequential module characterize user preferences from two different perspectives—spatial dependence and behavioral patterns—their representation distributions differ. Directly concatenating and fusing them may lead to insufficient alignment of cross-modal information. Therefore, this invention introduces a consistency learning framework to constrain the distribution consistency of the geographic representation and the sequential representation in the implicit representation space, thereby enhancing the complementarity and fusion effect of the two types of preference representations.

[0079] During the training phase, a memory bank is set up to store historical embeddings from different users.

[0080] For users Calculate the geographical representation separately With sequence representation And based on temperature-scaled cosine similarity and anchor representations in the memory, the components of the similarity distribution between the geographic module and the sequence module are obtained: , .

[0081] in, For cosine similarity, For temperature parameters, This represents the number of anchor points in the memory bank. and They represent the first in the memory bank. The geographic representation and sequence representation corresponding to each anchor point; by calculating the above... Each component constitutes a similarity distribution between the geographic module and the sequence module in the memory bank anchor point space. and : ,

[0082] To ensure that the geographic module and the sequence module learn a consistent distribution Introducing KL divergence to define consistency loss .

[0083] At the same time, the predicted probability of a user visiting a target POI is calculated using a multilayer perceptron. ;in, For the Sigmod function, Embedding of the target POI.

[0084] The supervised loss is calculated using binary cross-entropy. .

[0085] in For the training sample set, Indicates user Did you actually visit POI? Positive samples are assigned a value of 1, and negative samples are assigned a value of 0. Finally, the overall objective loss function of joint learning is obtained. .

[0086] in The weight hyperparameter is used to balance recommendation accuracy and consistency constraints. Through joint optimization, the model can enhance the complementarity of spatial and sequential features to improve the robustness and generalization ability of cross-regional recommendations.

[0087] The next-point-of-interest (POI) recommendation model proposed in this invention employs a geographic module with a graph neural network architecture to model spatial preferences based on spatial proximity relationships and cross-regional accessibility in the knowledge graph fused in step 2, extracting user spatial behavior features at different spatial scales. Simultaneously, a sequence module with a graph neural network architecture is used to model sequence preferences based on user access sequences depicted in the fused knowledge graph of step 2, learning long-term user preferences and higher-order travel patterns through attention mechanisms and random walk structures. Based on this, preference fusion prediction is performed. For the Next-POI user mobility behavior prediction task, spatial preference representation and sequence preference representation are jointly modeled through a fusion mechanism to generate the probability distribution of user access to candidate POIs, thus achieving Next-POI recommendation. The next-point-of-interest recommendation model in this invention has good scalability and transferability, and can be applied to various scenarios such as urban travel navigation, point-of-interest recommendation, smart tourism, and mobile services.

[0088] It should be noted that, in addition to using spatial object filtering based on OpenStreetMap, geospatial entity extraction methods can also use government open data, commercial map data, or POI data obtained from remote sensing image recognition.

[0089] In addition to using the regional heat matrix, the construction of cross-regional relationship weights can also use regional influence based on PageRank, travel probability based on the OD (Origin-Destination) matrix, or regional interaction strength modeling based on Poisson or Gaussian processes.

[0090] The weight scaling method can be replaced by the maximum relation weight method, or by traditional methods such as softmax normalization, min-max normalization, or weight regression models based on adaptive learning.

[0091] In addition to using MLP to output access probabilities, prediction can also employ logistic regression, LightGBM, graph convolution-based classifiers, dual-tower structures, or probabilistic prediction models based on energy functions.

[0092] In addition, to verify the effectiveness of the method of the present invention, a recommended quality assessment for Next-POI is also provided.

[0093] This invention systematically evaluates the quality of Next-POI recommendations from three perspectives: overall recommendation performance, performance in data sparsity and cold start scenarios, and the contributions of model components. The specific tasks include the following: a. Overall performance: The overall accuracy and loss metrics of the model were evaluated by comparing different cities and multiple baseline models.

[0094] b. Sparse scenes: Experiments with different proportions of training data were conducted to verify the model's adaptability to data sparsity and cold start scenarios.

[0095] c. Module contributions: The impact of geographic modules and cross-regional relationships on prediction performance was analyzed through ablation experiments.

[0096] To achieve the above objectives, this invention employs a multi-dimensional performance evaluation method based on unified data partitioning, model comparison experiments, and a modular ablation mechanism. The specific technical steps are as follows: First, the overall performance advantage is verified by comparing models from multiple cities and multiple baselines. Second, the stability of the model under sparse and cold-start scenarios is evaluated under different proportions of training data. Finally, the key role of the geographic module and cross-regional relationships in improving accuracy is analyzed through ablation experiments, providing a comprehensive quantitative basis for the reliability and generalization ability of the model.

[0097] I. Overall Performance Evaluation. Step I trains and tests the model under complete data conditions, and calculates AUC and Logloss to recommend performance metrics based on real check-in datasets from Tokyo and New York.

[0098] We compared our model with various graph neural network models and non-graph models to ensure that the experimental conditions were consistent.

[0099] Experiments have shown that, on the Tokyo dataset, the model of this invention achieves an AUC of 0.9465 and a Logloss of 0.3013, which is significantly better than the best baseline model LIMP (AUC 0.9209, Logloss 0.3733) and also better than the graph neural network baseline model MobGT (AUC 0.9114, Logloss 0.3951).

[0100] On the New York dataset, the model of this invention has an AUC of 0.9274 and a Logloss of 0.3496, which also comprehensively outperforms all the control models (MobGT has an AUC of 0.8943 and LIMP has an AUC of 0.9077).

[0101] The above results demonstrate that, in terms of overall recommendation accuracy, the method of this invention achieves state-of-the-art performance in both Tokyo and New York, thus enabling the formation of benchmark results with cross-regional stability.

[0102] II. Cold start and sparsity assessment. To verify the model's performance under conditions of insufficient training data, step II splits the training set into five proportions: 20%, 40%, 60%, 80%, and 100%, and trains the model independently under each proportion.

[0103] The performance of each model in the cold start scenario is shown in the appendix. Figure 3 As shown. Among them. Figure 3 (a), (b), (c), and (d) represent schematic diagrams of the cold start experiments for Tokyo AUC, New York AUC, Tokyo Logloss, and New York Logloss, respectively.

[0104] On the Tokyo dataset, even when training with only 20% of the data, the AUC remains at 0.8944 and the Logloss is 0.4361; when the training ratio is increased to 40%, the AUC exceeds 0.9, approaching the performance of training with the complete data.

[0105] A similar trend was observed in the New York dataset: the AUC was still 0.8735 when training with 20% of the data, while it was close to 0.9274 when training with 60% of the data.

[0106] The results show that even in highly sparse, data-limited, or user-cold-start environments, this invention maintains high prediction accuracy and stability without serious performance collapse, and possesses the generalization capabilities required for deployment in real-world urban services.

[0107] III. Evaluation of the contribution of model components.

[0108] To verify the contribution of geographic modules and cross-regional relationships to the final prediction performance, this step constructs three model variants: CRGNN with geographic modules removed w / o Geo, CRGNN with cross-regional relationships removed w / o Cross, and the complete model CRGNN.

[0109] Figure 4 A schematic diagram comparing the performance of various models in the ablation experiment is shown, in which... Figure 4 (a), (b), (c), and (d) represent schematic diagrams of ablation experiments in Tokyo AUC, New York AUC, Tokyo Logloss, and New York Logloss, respectively.

[0110] On the Tokyo dataset, removing the geo module decreased the AUC from 0.9465 to 0.8835, while the Logloss increased from 0.3013 to 0.4862. Removing cross-regional relationships resulted in an AUC of 0.9240, still significantly lower than the full model. On the New York dataset, the full model of this invention had an AUC of 0.9274, while the AUCs with and without Geo decreased to 0.8489 and 0.8932, respectively.

[0111] These results fully demonstrate that the geographic module is a fundamental component, and cross-regional relationships play a crucial role in improving the model's performance in regional boundaries and cross-regional travel scenarios. Both are indispensable structures for achieving high-precision recommendations in this invention.

[0112] Through steps I to III above, this invention systematically verifies the reliability, stability, and cross-regional generalization ability of the model from three perspectives: overall accuracy, data sparsity performance, and contribution of key modules.

[0113] Experimental results show that the model of this invention maintains industry-leading performance under different cities, different data scales and different structural configurations, and thus can meet the practical application needs of large-scale urban spatial service scenarios.

[0114] Compared with traditional point-of-interest recommendation methods, this invention differs significantly in at least three aspects: I. In terms of geospatial knowledge graph construction and cross-regional dynamic relationship modeling, a scalable geospatial knowledge graph model is constructed for massive urban spatial data. This model supports the unified extraction, cleaning, and structured representation of various POI spatial objects and administrative divisions, and automatically generates entity centroid coordinates, attribute labels, and multi-scale spatial proximity relationships. Cross-regional boundary identification is achieved based on administrative region filtering, distance threshold determination, and spatial distribution characteristics. Furthermore, an inter-regional heat matrix and user continuous check-in sequences are introduced to dynamically assign and update weights to cross-regional edges, enabling differentiated modeling of region-to-region migration intensity. This provides a precise and updatable spatial structural foundation for cross-regional knowledge fusion and user behavior modeling.

[0115] II. Regarding the construction of user preference knowledge graphs and the fusion of cross-regional knowledge graphs, a user preference knowledge graph is constructed using user check-in data. The geospatial knowledge graph and the user preference knowledge graph are aligned at the node and edge levels. Redundant entities and invalid edges are automatically removed while preserving key spatial relationships, achieving unified modeling of spatial structure and user behavior preferences. The fused cross-regional knowledge graph simultaneously includes proximity relationships, cross-regional relationships, and user behavior sequence relationships, forming a structurally consistent and semantically complete multi-relational city knowledge representation, providing a unified input carrier for subsequent GNN models.

[0116] III. In predicting Next-POI user mobility behavior, a graph neural network is employed to jointly learn spatial adjacency, cross-regional accessibility, and sequential behavior patterns. Through attention mechanisms and random walk structures, long-term user preferences and high-order travel patterns are captured. This function supports real-time calculation and ranking of user access probabilities, possesses scalability and transferability, and enables Next-POI prediction. It can be used in various scenarios such as urban travel navigation, point-of-interest recommendation, smart tourism, and mobile services.

[0117] This invention utilizes real-world urban spatial data, administrative division data, and user check-in behavior data for modeling. It can be applied in practical scenarios such as smart mobility, urban services, and cultural tourism recommendations. Its data sources and processing methods are closer to real business needs, which helps the model maintain stable performance during actual deployment. Furthermore, this invention can automatically generate spatial relationships, dynamic weights across regions, and complete graph fusion at the scale of millions of POIs and hundreds of administrative regions. Through graph neural networks, it achieves joint spatial-behavioral modeling, resulting in recommendation performance significantly superior to traditional models, thus possessing high-precision and high-efficiency practical recommendation capabilities. Moreover, this invention demonstrates good predictive performance even with sparse training data or insufficient user behavior data, maintaining stable recommendation performance without significant performance degradation. Compared to traditional recommendation methods that rely on depth of behavioral sequences, it has stronger robustness and generalization ability, making it suitable for various use cases in real-world urban services.

[0118] Example 2 This embodiment 2 describes a next point of interest recommendation system based on a cross-regional urban spatial knowledge graph, which is based on the same inventive concept as the next point of interest recommendation method based on a cross-regional urban spatial knowledge graph in embodiment 1 above.

[0119] The next point of interest recommendation system based on a cross-regional urban spatial knowledge graph in this embodiment includes the following modules: The module for constructing geospatial knowledge graphs and modeling cross-regional dynamic relationships is used to build scalable geospatial knowledge graphs for massive urban spatial data; it performs spatial data cleaning by combining administrative division data and user check-in data; on this basis, it generates spatial relationship structures between points of interest and constructs spatial edges that include spatial proximity relationships and cross-regional relationships. A unified fixed weight is used to model spatial proximity relationships, while an inter-regional heat matrix is ​​introduced for cross-regional spatial relationships. Differentiated weights are assigned to cross-regional edges, and dynamic updates of the weights are supported to characterize the differences in migration intensity between regions. The user preference knowledge graph construction and cross-regional knowledge graph fusion module is used to construct a sequence graph based on the time sequence of user check-in data, generate a user preference knowledge graph that describes user behavior preferences and migration patterns, and carry out cross-regional knowledge graph fusion in combination with the constructed geospatial knowledge graph. And a next point of interest recommendation module, used to build a next point of interest recommendation model that includes a geographic module, a sequence module, and a preference fusion prediction module; the geographic module and the sequence module are both implemented based on a graph neural network architecture; The geography module uses graph neural networks to model spatial preferences for spatial proximity relationships and cross-regional accessibility in the fused knowledge graph, in order to extract users' spatial behavior characteristics at different spatial scales; The sequence module performs sequence preference modeling based on the user access sequence depicted in the fused knowledge graph. It learns long-term user preferences and high-order travel patterns through attention mechanisms and random walk structures to perform preference fusion prediction. The preference fusion prediction module jointly models spatial preference representation and sequential preference representation through a fusion mechanism to generate the probability distribution of user access to potential points of interest, thereby enabling the recommendation of the next point of interest in the user's mobile behavior prediction task.

[0120] It should be noted that any content not mentioned in the above-described functional modules of the system described in Embodiment 2 can be referred to the step description of the corresponding method in Embodiment 1 above, and will not be repeated in detail here.

[0121] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A method for recommending next point of interest based on cross-regional urban spatial knowledge graphs, characterized in that, Includes the following steps: Step 1. Construct a scalable geospatial knowledge graph based on massive urban spatial data; perform spatial data cleaning by combining administrative division data and user check-in data; and generate a spatial relationship structure between points of interest based on this data. The spatial relationship structure includes spatial edges of spatial proximity relationships and cross-regional relationships. Spatial proximity relationships are modeled using a unified fixed weight, while cross-regional spatial relationships are modeled using an inter-regional heat matrix. Differentiated weights are assigned to cross-regional edges, and dynamic updates of the weights are supported to characterize the differences in migration intensity between regions. Step 2. Construct a sequence graph based on the time sequence of user check-in data to generate a user preference knowledge graph that characterizes user behavior preferences and migration patterns, and combine it with the constructed geospatial knowledge graph to carry out cross-regional knowledge graph fusion; Step 3. Build a next point of interest recommendation model that includes a geographic module, a sequence module, and a preference fusion prediction module; the geographic module and the sequence module are both implemented based on a graph neural network architecture; The geography module uses graph neural networks to model spatial preferences for spatial proximity relationships and cross-regional accessibility in the fused knowledge graph, in order to extract users' spatial behavior characteristics at different spatial scales; The sequence module performs sequence preference modeling based on the user access sequence depicted in the fused knowledge graph. It learns long-term user preferences and high-order travel patterns through attention mechanisms and random walk structures to perform preference fusion prediction. The preference fusion prediction module jointly models spatial preference representation and sequential preference representation through a fusion mechanism to generate the probability distribution of user access to potential points of interest, thereby enabling the recommendation of the next point of interest in the user's mobile behavior prediction task.

2. The next point of interest recommendation method based on cross-regional urban spatial knowledge graph according to claim 1, characterized in that, In step 1, spatial data cleaning is performed by combining administrative division data and user check-in data. Only administrative regions with user check-in behavior are retained, and invalid or redundant spatial objects are removed. The process is as follows: Construct a region set based on administrative boundary data. By combining user check-in data, each administrative region can be monitored. Whether or not a valid check-in point is included, only administrative regions with at least one check-in point are retained, and all other regions are considered invalid and are removed; Filtered set of administrative regions ;in Indicates points of interest in geographic space. This represents the set of points of interest in the user check-in dataset. Indicates the use of points of interest. The region to which it belongs; For raw objects in the OpenStreetMap data source Only spatial objects located within the effective administrative area are retained. And remove objects with missing coordinates, missing attributes, or no name; Generate a uniform format for the preserved spatial objects, including the name field. Centroid latitude and longitude coordinates Functional attribute tags Used for subsequent spatial relationship construction; where the centroid coordinates Calculated using the following formula: ； in Number of boundary points: If it is a point object, =1; if it is a road object or a region object =Number of outline points of the object; , These represent the first and second parts of the object. The latitude and longitude of each boundary point; The cleaned POI data is retained and denoted as a geospatial knowledge graph. POI set .

3. The next point of interest recommendation method based on cross-regional urban spatial knowledge graph according to claim 1, characterized in that, In step 1, the process of generating spatial proximity relationships and cross-regional relationships is as follows: For any two POIs, define them respectively as follows: , The spatial relationship between the two can be determined in the following way: ； in express and Spatial relationships between them; This represents a spatial proximity edge generated within the same region. Indicates a cross-regional relationship edge; pass The function checks if objects are in the same domain. and Within the same domain and at a distance less than or equal to the nearest neighbor threshold If they are spatially adjacent, then they are spatially adjacent; if and Foreign regions and distance less than or equal to the cross-region threshold This indicates a cross-regional relationship; All eligible spatial proximity relationships and cross-regional relationships will be added to the spatial relationship set. ; This leads to the formation of a geospatial knowledge graph. .

4. The next point of interest recommendation method based on cross-regional urban spatial knowledge graph according to claim 1, characterized in that, In step 1, the process of setting spatial relationship weights is as follows: Spatial proximity edges generated within the same region Assign a fixed weight constant of 1 to all adjacent edges; For cross-regional relationship edges The weights are dynamically allocated based on the regional popularity. Based on users' consecutive check-in sequences, cross-regional access patterns across all regions are statistically analyzed to construct a regional popularity matrix. : ； in For the number of regions, This indicates that the user's consecutive check-ins are from the region. To the region The number of cross-regional behaviors; Cross-regional relationship weight Defined as: ； in For the region The maximum relation weight is the global scaling factor; Indicates the area The target region with the highest popularity among all cross-regional activities; Through scaling, weight Limited to Inside; Specifically, the maximum relation weight for: ； in Indicates from the region Number of cross-regional activities to other regions; Indicates all regions, from region The maximum number of cross-regional trips to other regions; this ratio reflects the region's... The value is taken from the perspective of the overall impact on travel. Within the range; Through the weight setting process, structurally consistent, behavior-driven, and dynamically adjustable edge weights are generated for all spatial relationships, thereby providing high-quality spatial input for subsequent cross-regional knowledge graph construction and graph neural network modeling.

5. The next point of interest recommendation method based on cross-regional urban spatial knowledge graph according to claim 1, characterized in that, In step 2, the process of constructing the user preference knowledge graph is as follows: For users Sign-in data Perform parsing and extract the user's information. Collection of POIs accessed And construct directed action edges for adjacent POIs that check in. ; in express The number of times you sign in. Indicates user No. The POI for the first check-in, and the two consecutive check-ins , There is a user preference relationship. Merge all relationships to obtain set of preference relations ; Thus, this user was formed. User Preference Knowledge Graph After integrating the preference paths of all users, a complete user preference knowledge graph is constructed. ; The set of nodes , indicating all users Points of Interest (POIs) for check-in and their order; edge set , indicating all users The order of visits between adjacent Points of Interest (POIs) in the check-in sequence.

6. The next point of interest recommendation method based on cross-regional urban spatial knowledge graph according to claim 1, characterized in that, In step 2, the cross-regional knowledge graph fusion performs alignment and merging at the node and edge levels, automatically eliminating redundant entities and invalid edges while retaining key spatial and behavioral relationships, thus forming a unified cross-regional knowledge graph representation. The fused knowledge graph simultaneously includes spatial proximity relationships, cross-regional relationships, and user behavior sequence relationships, constructing a structurally consistent and semantically complete multi-relational urban knowledge representation. The process of cross-regional knowledge graph fusion is as follows: Based on the geospatial knowledge graph, node mapping is performed on the user preference knowledge graph; For any POI in the user preference knowledge graph, it is defined as Node set Represent all user check-in POIs and their order; find the nearest node in the geospatial knowledge graph. , ; in Represents the geographical distance between any two POIs; through Choose the node with the smallest distance as the mapped node; Representing geospatial knowledge graphs The set of POIs; After processing the node mapping, the user preference knowledge graph edge set is... Each edge in Mapped to Add it to the new edge set In, retain the appearance of The nodes in the graph are used to obtain the node set of the merged knowledge graph: ； Simultaneously, node sets are selected from the geospatial knowledge graph. The spatial relationships between them yield the preserved set of geographical edges. This removes invalid nodes and edges. This is a set of spatial relationships consisting of all eligible spatial proximity and cross-regional relationships. Finally, the relationship between the geospatial knowledge graph and the user preference knowledge graph is merged to obtain the fused knowledge graph. The merged knowledge graph It possesses both geospatial structure and behavioral sequence information.

7. The next point of interest recommendation method based on cross-regional urban spatial knowledge graph according to claim 1, characterized in that, In step 3, the processing flow of the geographic module is as follows: In a geospatial graph neural network, each POI node The layer Represented as: ； in, For this node First-order adjacency set A node in; For nodes to its adjacent nodes The vector for message passing; For the first Layer trainable weight matrix; ； in Indicates the weight of cross-regional relationships. Indicates the distance between two nodes; go through Layered message passing obtains all POI geographic representations. : ； For users Sign-in sequence Extract the visited POIs to obtain a geographic representation. ; Utilizing multi-head attention mechanism for users Behavioral preferences are aggregated to obtain their semantic representation. : ； in, , ; , , The first Attention header query, key-value mapping matrix For the original embedding dimension, For the number of attention heads, For POI In the The attention score in each self-attention head, i.e., the self-attention weight; Depend on Perform queries and key-value mappings, then dot products, and finally scale the results; Softmax normalization is performed to obtain .

8. The next point of interest recommendation method based on cross-regional urban spatial knowledge graph according to claim 1, characterized in that, In step 3, the processing procedure of the sequence module is as follows: Given user User Preference Knowledge Graph ; The set of nodes The set of edges represents the points of interest (POIs) visited by the user. This indicates a directed transition relationship between adjacent sign-ins; For any two graphs and Its direct product graph is defined as ; in , , , The diagrams are shown below. , The set of nodes and edges of a , the set of nodes in a , or the set of edges in a , or ... edge set ; exist A random walk on is defined as simultaneously on and Synchronous walk on; assuming for adjacency matrix If it is a vector of all 1s, then Step random walk kernel Used to measure the length of two graphs. Path similarity; in In the product graph The upper length is The counting matrix of all random walks; There is a set Image filter And for each user User Preference Knowledge Graph Kernel similarity calculation is performed with the graph filter; where, This represents a set of trainable graph filters; For the set of random walk step sizes The characteristic matrix is ​​obtained. , The Middle The row corresponds to the first The similarity vector of the graph filter, the first Column corresponding random walk step size ; Feature matrix matrix elements Defined as: ； matrix Flatten and stitch together before inputting into a fully connected layer. Get users Sequence semantic representation ; For splicing operations, This represents a higher-order behavioral preference vector for users in the sequence dimension.

9. The next point of interest recommendation method based on cross-regional urban spatial knowledge graph according to claim 1, characterized in that, In step 3, a consistency learning framework is introduced to constrain the distribution consistency of the two representations in the implicit space. During the training phase, a memory bank is set up to store historical embeddings from different users; For users Calculate the geographical representation separately With sequence representation And based on temperature-scaled cosine similarity and anchor representations in the memory, the components of the similarity distribution between the geographic module and the sequence module are obtained: 、 ； in, For cosine similarity, For temperature parameters, This represents the number of anchor points in the memory bank. and They represent the first in the memory bank. The geographic representation and sequence representation corresponding to each anchor point; by calculating the above... Each component constitutes a similarity distribution between the geographic module and the sequence module in the memory bank anchor point space. and : 、 To ensure that the geographic module and the sequence module learn a consistent distribution Introducing KL divergence to define consistency loss ; At the same time, the predicted probability of a user visiting a target POI is calculated using a multilayer perceptron. ;in, For the Sigmod function, Embedding of the target POI; The supervised loss is calculated using binary cross-entropy. ; in For the training sample set, Indicates user Did you actually visit POI? Positive samples are assigned a value of 1, and negative samples are assigned a value of 0. Finally, the overall objective loss function for joint learning is obtained. ; in The weight hyperparameter is used to balance recommendation accuracy and consistency constraints. Through joint optimization, the model can enhance the complementarity of spatial and sequential features to improve the robustness and generalization ability of cross-regional recommendations.

10. A next point of interest recommendation system based on cross-regional urban spatial knowledge graph, characterized in that, Includes the following modules: The module for constructing geospatial knowledge graphs and modeling cross-regional dynamic relationships is used to build scalable geospatial knowledge graphs for massive urban spatial data; it performs spatial data cleaning by combining administrative division data and user check-in data; on this basis, it generates spatial relationship structures between points of interest and constructs spatial edges that include spatial proximity relationships and cross-regional relationships. A unified fixed weight is used to model spatial proximity relationships, while an inter-regional heat matrix is ​​introduced for cross-regional spatial relationships. Differentiated weights are assigned to cross-regional edges, and dynamic updates of the weights are supported to characterize the differences in migration intensity between regions. The user preference knowledge graph construction and cross-regional knowledge graph fusion module is used to construct a sequence graph based on the time sequence of user check-in data, generate a user preference knowledge graph that describes user behavior preferences and migration patterns, and carry out cross-regional knowledge graph fusion in combination with the constructed geospatial knowledge graph. And a next point of interest recommendation module, used to build a next point of interest recommendation model that includes a geographic module, a sequence module, and a preference fusion prediction module; the geographic module and the sequence module are both implemented based on a graph neural network architecture; The geography module uses graph neural networks to model spatial preferences for spatial proximity relationships and cross-regional accessibility in the fused knowledge graph, in order to extract users' spatial behavior characteristics at different spatial scales; The sequence module performs sequence preference modeling based on the user access sequence depicted in the fused knowledge graph. It learns long-term user preferences and high-order travel patterns through attention mechanisms and random walk structures to perform preference fusion prediction. The preference fusion prediction module jointly models spatial preference representation and sequential preference representation through a fusion mechanism to generate the probability distribution of user access to potential points of interest, thereby enabling the recommendation of the next point of interest in the user's mobile behavior prediction task.

Images