Unsupervised region representation learning method and system for learning urban region embedding

Abstract

The invention relates to an unsupervised region representation learning method and system for learning urban region embedding, and belongs to the technical field of urban calculation and intelligent systems. The method specifically comprises the following steps: S1, constructing a regional mixed graph network, and capturing pairing and grouping relationships among regions by using multi-modal data; s2, graph and hypergraph comparative learning is executed, and discriminative representation of region nodes and a high-order relation between regions are learned from graph and hypergraph structures through a parallel comparative learning module; s3, cross-module comparative learning is carried out, information exchange between graph and hypergraph node representation is promoted through a controller assembly, and the learning ability of the model is enhanced; and S4, optimizing model parameters, and generating more effective region embedding through joint optimization of graph loss, hypergraph loss and cross-module loss so as to support various downstream tasks. According to the technical scheme provided by the invention, the accuracy and efficiency of urban area analysis are improved, and the capturing capability of the model on complex interaction between the areas is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of urban computing and intelligent systems, and relates to an unsupervised region representation learning method and system for learning urban region embedding. Background Art

[0002] With the acceleration of urbanization and the development of information technology, the explosive growth of urban data has provided unprecedented opportunities and challenges for many fields such as urban planning, traffic management, and public safety. Urban regional analysis has become increasingly important in the fields of urban planning, traffic management, and public safety. Traditional urban regional analysis methods are often limited to independently evaluating the characteristics and patterns of each region, ignoring the complex interactions and dynamic relationships between regions, which greatly limits the accuracy and practicality of the analysis results.

[0003] The complex relationship between urban areas is reflected in many aspects. First, the geographical proximity makes different areas closely connected in terms of human flow, logistics, information flow, etc. For example, the human flow attraction of a commercial center is not only affected by its own business atmosphere and the degree of facility perfection, but also closely related to factors such as the convenience of surrounding transportation and residential population density. Similarly, housing prices in residential areas are also affected by multiple factors such as surrounding educational resources, medical services, and green environment.

[0004] Secondly, the functional diversity and spatial heterogeneity of urban areas further increase the difficulty of analysis. Different areas may carry multiple functions such as residence, commerce, industry, culture and entertainment, and each functional area has its own unique operating rules and influencing factors. Therefore, when conducting urban regional analysis, it is necessary to comprehensively consider multiple factors such as regional functions, land use conditions, population density, and economic development level to ensure the comprehensiveness and accuracy of the analysis results.

[0005] However, traditional urban regional analysis methods are insufficient in dealing with these complex problems. On the one hand, the information provided by a single data source is often too one-sided and cannot fully reflect the real situation of the region; on the other hand, simple feature engineering and statistical models are difficult to capture the complex relationships and dynamic changes between regions. Therefore, researchers began to seek new methods to meet these challenges.

[0006] In recent years, the rise of deep learning technology has provided new ideas for urban regional analysis. Deep learning models have powerful feature extraction and pattern recognition capabilities, and can automatically learn useful information from large-scale data without too much human intervention. By applying deep learning technology to the field of urban regional analysis, more complex and more accurate models can be built to capture the complex relationships and dynamic changes between regions.

[0007] However, applying deep learning technology to urban regional analysis also faces many challenges. First, how to effectively collect and integrate multi-source heterogeneous urban data is a key issue. Different data sources differ in format, accuracy, timeliness, etc., and appropriate data preprocessing and fusion technologies are needed to ensure data consistency and availability. Secondly, how to design a reasonable model structure and parameters to adapt to the characteristics of different cities and regions is also an important issue. Different cities differ in geographical environment, cultural background, economic development, etc., and models need to be optimized and adjusted according to specific circumstances.

[0008] In summary, there is an urgent need for a universal method that can capture the complex relationships between urban areas for effective urban area analysis. This method should be able to process large-scale, multimodal urban data and automatically learn useful information from the data to guide decision-making in areas such as urban planning, traffic management, and public safety. At the same time, the method also needs to be able to adapt to different urban and regional characteristics and provide accurate and robust regional embedding results. Summary of the invention

[0009] In view of this, the purpose of the present invention is to provide an unsupervised region representation learning method and system for learning urban region embedding, which learns comprehensive region embedding from multimodal data by integrating graph and hypergraph comparative learning, can capture pairwise and high-order relationships between regions, and automatically adapt to different urban environments and task requirements. The technical solution provided by the present invention can achieve a deeper understanding of urban regions and provide more scientific and effective support for urban region analysis.

[0010] In order to achieve the above object, the present invention provides the following technical solutions:

[0011] An unsupervised region representation learning method for learning urban region embeddings, the method comprising the following steps:

[0012] S1. Construct a regional hybrid graph network model to capture the pairwise and grouped relationships between regions using multimodal data.

[0013] S2, perform graph and hypergraph contrastive learning, through parallel contrastive learning modules, learn the discriminative representation of regional nodes and high-order relations between regions from the graph and hypergraph structures respectively;

[0014] S3, cross-module comparative learning, promotes information exchange between graph and hypergraph node representations through controller components, and enhances the learning ability of the model;

[0015] S4. Optimize model parameters by jointly optimizing graph loss, hypergraph loss, and cross-module loss to generate more effective region embeddings to support various downstream tasks.

[0016] Furthermore, in step S1, a street view context embedding step and a paired edge construction step are included, and the specific steps are as follows:

[0017] S11, Street view context embedding: Using street view image data, the visual encoder extracts the visual features in the area and encodes these features into the initial node embedding of the area; the Inception-V3 architecture is used, and the pre-trained weights are imported as the encoder of the street view context embedding; for a given area r i A set of street view images The final visual embedding where f η represents the visual encoder, a represents r i The number of randomly sampled street view images in ; the embedding of all regions is represented as These embeddings capture the physical environment and visual information in the area, providing rich contextual information for subsequent graph learning; Inception-V3 is a deep convolutional neural network developed by Google. It is an improved version of the earlier version of the Inception series. Inception-V3 is widely used in image classification, target detection and other tasks.

[0018] Google Inception Net won the first place in the 2014 ImageNet Large Scale Visual Recognition Competition (ILSVRC). The network won with structural innovation. By replacing the fully connected layer with a global average pooling layer, the number of parameters was greatly reduced. This network model is generally called Inception V1. In the subsequent Inception V2, the Batch Normalization method was introduced to accelerate the convergence of training. In the Inception V3 model, by splitting the two-dimensional convolution layer into two one-dimensional convolution layers, not only the number of parameters was reduced, but also the overfitting phenomenon was alleviated. Inception-V3 consists of an initial feature extraction module (Stem) and multiple Inception modules. Each module contains parallel 1x1, 3x3, 5x5 convolution and pooling operations to capture features of different scales. The convolution factorization technology decomposes a large convolution kernel into multiple small convolution kernels to reduce the amount of calculation. At the end of the network, global average pooling is used instead of the fully connected layer, and finally the classification result is output through the softmax layer. The entire architecture design focuses on efficiency and parameter optimization, which is suitable for large-scale image classification tasks.

[0019] S12. Paired edge construction: construct paired edges in a hybrid graph based on geographic proximity between regions and human mobility data; construct a matrix based on human mobility Represents region r i and r j The standardized number of trips between them is constructed based on the adjacency matrix A of the geographical neighbors. s ∈{0, 1} n×n ; In the initial edge matrix A′ f In the example, the KNN algorithm is used to retain the most relevant k1 edges (k1=10) of each node, and the updated edge matrix A is obtained. f , then, the fused adjacency matrix A = A is obtained by element-by-element logical OR operation s o A f , element a in matrix A ij Value represents area r i and r j Whether there are link edges between them, these edges represent the direct interaction and spatial relationship between regions, and the map of the regions is obtained.

[0020] S13, Hyperedge construction: To capture high-order relationships between regions, such as common POI distribution patterns or similar human mobility patterns, hyperedges are constructed to group multiple regional nodes together. First, based on human mobility, from the perspectives of the region’s source and destination, k2 (k2=10) regions closest to each region are selected to form hyperedges. All these hyperedge sets constitute a hyperedge association matrix based on mobility data. Secondly, by calculating the cosine similarity between POI features and selecting the most similar k3 (k3 = 10) nodes to form a hyperedge, the association matrix is ​​constructed similarly Then merge H by concatenation operation f and H p Generate the final hyperedge association matrix H = H f ||H p ,in m=m1+m2; get the hypergraph The introduction of hyperedges enables hybrid graphs to represent more complex relationships between regions.

[0021] S14, hybrid graph construction: While completing the street view context embedding and edge construction process, combine the edge set obtained in the previous steps and hyperedge sets Get based on region The mixed graph of is taken as the input of the framework, and the mixed graph is represented as

[0022] Furthermore, in step S2, three steps are included: graph / hypergraph enhancement, graph / hypergraph representation, and graph / hypergraph contrast loss function, specifically including:

[0023] S21. Graph / hypergraph enhancement: Use data augmentation techniques to perturb graph and hypergraph structures and generate different views to simulate and alleviate noise problems in graphs and hypergraphs. For graph structures, a random masking matrix is ​​sampled from a Bernoulli distribution with probability p. m Randomly mask some dimensions of the node feature matrix X with probability p r Randomly delete graph Some edges in the original graph Generate enhanced two-graph view and Similarly, for the hypergraph structure, with probability p′ m Randomly mask node features in different dimensions with probability p′ r Random deletion For some hyperedges in , we get two enhanced hypergraph views and Since nodes in graphs and hypergraphs essentially represent regions, feature masking is the same for both, i.e.

[0024] S22. Graph / hypergraph representation: Use graph convolutional networks (GCNs) and hypergraph neural networks (HGNNs) as encoders to learn representations of nodes and hyperedges from enhanced graphs and hypergraphs. Graph representation learning focuses on updating the embedding of each node by aggregating neighbor node information. and The input is fed into the shared parameter graph convolutional network GCN encoder to obtain the node embedding matrices Y1 and Y2 of the two views, and then the node embedding matrix is ​​projected into a smaller potential space through a two-layer MLP projection head to obtain the projected embedding matrix and Similarly, hypergraph representation learning uses a hypergraph neural network (HGNN) encoder to learn node and hyperedge representations on the two generated enhanced views. The node and hyperedge embedding matrices and trainable weight matrices are used to update the embeddings of nodes and hyperedges in an iterative manner. Finally, the node embedding matrices P1 and P2 and the hyperedge embedding matrices Q1 and Q2 of the two views are obtained. In order to improve the effect of contrastive learning, a two-layer MLP is used as a projection head to project the embeddings of nodes and hyperedges into smaller latent spaces, respectively, to obtain the embedding matrices of nodes and hyperedges of the corresponding views. and

[0025] S23. Graph / hypergraph contrastive loss: Design a contrastive loss function to guide the model to learn to distinguish the representations of the same region between different views and to bring the representations of similar regions in different views closer. For graph contrastive learning, use the InfoNCE loss function in contrastive learning to calculate the node-level contrastive loss. For any node v i, embedded in the first view As an anchor, the corresponding embedding in the second view As a positive sample, other nodes are used as negative samples. The node-level contrast loss is calculated by maximizing the similarity between positive samples (calculated by the cosine similarity function) and minimizing the similarity between the anchor point and the negative samples, that is, in τ v represents the temperature parameter of the hard negative sample penalty, s(·) is the cosine similarity function, which is used to measure the similarity between two representations. Similarly, for hypergraph contrastive learning, node-level, hyperedge-level, and membership-level contrastive losses are designed to capture high-order relationships in the hypergraph and promote the learning of representations with high-order information. After obtaining the projected node and hyperedge embedding matrices, the node-level, hyperedge-level, and membership-level contrastive losses are calculated respectively. The node-level contrastive loss is calculated by maximizing the representation similarity of the same node in different views while minimizing the similarity between different nodes, i.e. in The hyperedge-level contrastive loss is calculated by maximizing the representation similarity of the same hyperedge in different views, i.e. in The contrast loss at the affiliation level considers the association between nodes and hyperedges and enhances the representation of nodes and hyperedges, i.e. in Finally, the weighted sum of the three levels of contrast loss is used to obtain the total hypergraph contrast loss

[0026] Furthermore, in step S3, the quality and expression ability of region embedding are further optimized by introducing a cross-module contrast learning mechanism; specifically, the following steps are involved:

[0027] S31. Information fusion: In order to integrate the two sets of region embedding representations obtained in the parallel graph contrastive learning module and the hypergraph contrastive learning module, a weight-based fusion mechanism is adopted to integrate the node embedding representations Z from different modules. v and Z n Perform weighted 2 fusion to obtain the fusion representation Z f =αZ v +(1-α)Z n , where α∈(0,1) is a weight parameter used to balance the contribution of different modules;

[0028] S32. Contrastive loss calculation: Using the fused node embedding representation Z f , the fusion contrast loss is defined by calculating the contrast loss between the same node pairs Then the contrast losses of all nodes are combined to obtain the overall loss function of cross-module contrastive learning To optimize the final node representation.

[0029] Furthermore, in step S4, a comprehensive region embedding learning process is implemented to obtain the final region embedding by jointly optimizing the loss functions of all contrastive learning modules; specifically, the process includes:

[0030] The overall loss is defined as Where λ1 and λ2 are the The parameters of importance; the final region representation is obtained after model training In each training cycle, the graph and hypergraph enhancement operations are performed, first from and Generate two enhanced graphs and hypergraphs, which are then fed into the corresponding shared graph encoders f θ (·) and hypergraph encoder to learn relevant representations; subsequently, different projection heads (p φ (·), p ψ (·)and ) maps these embeddings into the latent space and calculates the contrastive loss of the graph and hypergraph modules respectively; next, a controller component is used to fuse the node representations of the graph and hypergraph modules, resulting in a cross-module contrastive loss; finally, the model is trained by optimizing the overall objective loss; after training, the well-trained model will output the region representation E, which is applied to various downstream tasks.

[0031] The present invention also provides an unsupervised region representation learning system for learning urban region embedding.

[0032] The beneficial effects of the present invention are:

[0033] The present invention provides a framework called HyperRegion, which is specially designed for urban computing tasks and learns comprehensive regional embedding from multimodal data in an unsupervised manner. The framework can capture pairwise and high-order relationships between regions by integrating graph and hypergraph contrastive learning, thereby providing a more comprehensive regional representation. The HyperRegion framework promotes information exchange and collaboration between different modules by executing graph contrastive learning and hypergraph contrastive learning in parallel, and introducing cross-module contrastive learning, which optimizes the model's ability to capture complex interactions between regions and greatly improves the quality and robustness of regional embedding. The framework not only improves the accuracy of regional representation, but also provides a highly adaptable solution for different urban regions and downstream tasks. In addition, the design of HyperRegion takes into account the noise and incompleteness problems in the data, and enhances the robustness and generalization ability of the learned representation through the self-supervised contrastive learning paradigm. In general, the present invention greatly improves the efficiency and effectiveness of urban regional analysis, which is of great significance to the future development of urban computing and intelligent systems.

[0034] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0036] Figure 1 A system block diagram of the solution of the present invention;

[0037] Figure 2 Flow chart of the scheme of the present invention. DETAILED DESCRIPTION

[0038] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0039] The present invention proposes an unsupervised region representation learning method and system for learning urban region embedding, and the method proposes a framework called HyperRegion. The framework first uses multimodal data and graph neural networks to capture key features inside and outside the region, which are considered to be crucial for determining regional embedding. Then, HyperRegion constructs a regional hybrid graph network by comprehensively considering human mobility, geographic proximity, point of interest distribution, and visual semantics. In this embodiment, the proposed HyperRegion framework is extensively evaluated in different urban environments (such as different urban areas and multimodal datasets) and multiple downstream tasks (such as crime prediction, check-in prediction, and land use classification). Empirical results show that HyperRegion shows significant effectiveness and efficiency in performance, effectively solving the limitations of existing methods in processing complex urban data.

[0040] Figure 1 is a system block diagram of the solution of the present invention, Figure 2 The following is a flow chart of the present invention. In the technical solution of the present invention, the seven concepts involved are as follows:

[0041] Concept 1: Points of Interest (POI): Points of interest are places in the city with specific functions or significance, such as restaurants, shops, schools, and hospitals. These places are characterized by the services or activities they provide and have an important impact on the daily lives of city residents. In the HyperRegion framework, the distribution and characteristics of POIs are used to enrich the region embedding and help understand the function and attractiveness of the region.

[0042] The second concept: human mobility: Human mobility refers to people's movement patterns and behaviors in the city, including commuting, leisure and other daily activities. This mobility not only reflects the interaction between urban regions, but also affects urban dynamics and regional vitality. In the HyperRegion framework, human mobility data is used to capture the dynamic connections between regions and provide a basis for urban transportation planning and regional development analysis.

[0043] The third concept: Geographic neighbors: Geographic neighbors describe the spatial proximity between regions, which is based on the proximity of geographical locations. The concept of geographic neighbors is very important when building graph structures because it directly affects how the edges between regions are connected. In the HyperRegion framework, geographic neighbor relationships are represented by paired edges, which helps to model local spatial relationships and interactions between regions.

[0044] The fourth concept: Street view images: Street view images provide visual information of urban areas, including physical environmental features such as buildings, roads, and public spaces. These image data can be used to extract visual features of the region, such as architectural style, greening level, and pedestrian density. In the HyperRegion framework, street view image data can be used to initialize the visual embedding of regional nodes, or to further enrich the multimodal feature representation of the region through a visual encoder.

[0045] The fifth concept: Regional Embedding: It involves converting regions in a city into a low-dimensional vector space. These embedding vectors can retain the geographical and functional information of the regions, as well as the relationships between regions, so that complex urban data can be effectively processed by machine learning algorithms. In regional embedding, the spatial and semantic relationships between regions are usually maintained. For example, if two regions are geographically adjacent or functionally similar, their vector representations should also be close to each other. In the HyperRegion framework, these embedding vectors are used to determine the characteristics of urban regions and provide support for urban analysis and prediction tasks. By learning these embeddings, the system can better understand and predict the role of each region in urban functions and activities, thereby making more accurate regional function identification and prediction.

[0046] Concept 6: Graph Contrastive Learning: Graph contrastive learning is a self-supervised learning method that learns discriminative embeddings by maximizing the consistency between graph representations under different views. In the HyperRegion framework, the graph contrastive learning module generates different graph views by applying graph augmentation strategies and designs contrastive loss functions to guide model learning. This approach can improve the robustness and discriminability of the learned region embeddings, especially in the presence of data noise and incompleteness.

[0047] Concept 7: Hypergraph contrastive learning: Hypergraph contrastive learning is an extension of graph contrastive learning, which naturally models high-order relationships between multiple entities by introducing hyperedges. In the HyperRegion framework, the hypergraph contrastive learning module not only considers the pairwise relationships between regions, but also captures the grouped relationships between regions, thus providing a more comprehensive description of regional interactions. This approach helps to learn regional representations that contain high-order information and enhances the model's ability to understand complex urban phenomena.

[0048] like Figure 1 and Figure 2 As shown, the system framework of the present invention mainly includes four modules: hybrid graph construction module, graph comparison learning module, hypergraph comparison learning module, and cross-module comparison module. Figure 1 The system block diagram of the present invention is shown, wherein: Hybrid graph construction module: This module is intended to integrate multimodal data sources and construct a regional hybrid graph network, specifically including the following steps:

[0049] Step 1: Using street view image data, extract visual features X in the region through a visual encoder and encode these features into initial node embeddings of the region. These embeddings capture the physical environment and visual information in the region, providing rich contextual information for subsequent graph learning;

[0050] Step 2: Based on the geographic proximity between regions and human mobility data, construct the pairwise edges in the hybrid graph. In the initial edge matrix A′ constructed based on human mobility f In the above example, the KNN algorithm is used to retain the most relevant k1 edges of each node, and the updated edge matrix A is obtained. f , then, construct the matrix A by combining the adjacency matrix based on geographic neighbors s The fused adjacency matrix A is obtained by element-by-element logical OR operation. These edges represent the direct interaction and spatial relationship between regions, and the region graph is obtained.

[0051] Step 3: In order to capture the high-order relationships between regions, such as common POI distribution patterns or similar human mobility patterns, hyperedges are constructed to group multiple regional nodes together. First, based on human mobility, from the perspectives of the source and destination of the region, the k2 regions closest to each region are selected to form hyperedges, forming a hyperedge association matrix H based on mobility data. f Secondly, by calculating the cosine similarity between POI features and selecting the most similar k3 nodes to form a hyperedge, the association matrix H is constructed similarly. p Then H is merged by a join operation f and H p Generate the final hyperedge association matrix H and get the hypergraph The introduction of hyperedges enables hybrid graphs to represent more complex relationships between regions.

[0052] Step 4: Integrate the attributes within the region and the relationships between regions through the graph and hypergraph structures to build a regional hybrid graph that contains graph and hypergraph structures Each node v represents a region, and the relationship between regions is represented by the edge and hyperedge connection matrices A and H, providing input for subsequent contrastive learning.

[0053] Graph contrast learning module: The purpose of this module is to learn the discriminative representation of regional nodes through graph structure, including the following steps:

[0054] Step 1: Apply graph enhancement strategies such as node feature masking and edge removal to generate two different graph views and

[0055] Step 2: Use a shared graph neural network based encoder f θ(·) Learn node representations from these two augmented views and use the projection head p φ (·) Map node representations to a low-dimensional latent space.

[0056] Step 3: Calculate contrast loss By optimizing this loss function, the model is encouraged to learn robust and discriminative node representations that can distinguish different views.

[0057] Hypergraph Contrastive Learning Module: This module aims to learn high-order relationships between regions through the hypergraph structure, including the following steps:

[0058] Step 1: Apply hypergraph enhancement strategies, such as node feature masking and hyperedge removal, to generate two different hypergraph views and

[0059] Step 2: Use a shared hypergraph neural network based encoder to form node and hyperedge representations, and use the projection head p ψ (·)and Map them into a low-dimensional latent space.

[0060] Step 3: Based on the InfoNCE loss function in contrastive learning, a three-level contrastive loss mechanism is designed, including node-level, hyper-edge-level, and membership contrastive losses. Finally, the three-level contrastive losses are weighted summed to obtain the total hypergraph contrastive loss. By optimizing these loss functions, the model is encouraged to learn representations with high-order information from the hypergraph structure.

[0061] Cross-module comparison module: In order to enhance the information exchange and representation complementarity between different modules, this module uses a controller component to fuse the node representations learned from the graph and hypergraph, including the following steps:

[0062] Step 1: Use the weighted fusion mechanism to fuse the node representations learned from the graph and hypergraph to form a fused representation Z f .

[0063] Step 2: Calculate fusion contrast loss This loss takes into account the representation of the same node in different modules, encouraging the model to integrate common and complementary information.

[0064] Step 3: Jointly optimize all contrast losses, including image contrast loss Hypergraph Contrastive Loss And cross-module comparison loss to generate more effective region embeddings.

[0065] In summary, this paper proposes a framework called HyperRegion, which first integrates multimodal data using a regional feature fusion module, and then learns the local and global features of the region through a graph contrast learning module and a hypergraph contrast learning module. Finally, the information exchange and complementarity of representations between different modules are enhanced through a cross-module contrast module. The proposed HyperRegion is extensively evaluated in different urban computing tasks (such as check-in prediction and crime prediction). Empirical results show that HyperRegion shows significant effectiveness and superiority in performance, effectively solving the limitations of existing methods in regional representation learning.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. An unsupervised region representation learning method for learning urban region embeddings, characterized by: The method comprises the following steps: S1. Construct a regional hybrid graph network model to capture the pairwise and grouped relationships between regions using multimodal data. S2, perform graph and hypergraph contrastive learning, through parallel contrastive learning modules, learn the discriminative representation of regional nodes and high-order relations between regions from the graph and hypergraph structures respectively; S3, cross-module comparative learning, promotes information exchange between graph and hypergraph node representations through controller components, and enhances the learning ability of the model; S4. Optimize model parameters by jointly optimizing graph loss, hypergraph loss, and cross-module loss to generate more effective region embeddings to support various downstream tasks.

2. The unsupervised region representation learning method for learning urban region embedding according to claim 1, characterized in that: In step S1, including the street view context embedding step and the paired edge construction step, the specific steps are as follows: S11, Street view context embedding: Using street view image data, the visual encoder extracts the visual features in the area and encodes these features into the initial node embedding of the area; the Inception-V3 architecture is used, and the pre-trained weights are imported as the encoder of the street view context embedding; for a given area r i A set of street view images The final visual embedding where f η represents the visual encoder, a represents r i The number of randomly sampled street view images in ; the embedding of all regions is represented as S12. Paired edge construction: construct paired edges in a hybrid graph based on geographic proximity between regions and human mobility data; construct a matrix based on human mobility Represents region r i and r j The standardized number of trips between them is constructed based on the adjacency matrix A of the geographical neighbors. s ∈{0,1} n×n ; In the initial edge matrix A' f In the example, the KNN algorithm is used to retain the most relevant k1 edges (k1=10) of each node, and the updated edge matrix A is obtained. f , then, the fused adjacency matrix is ​​obtained by element-by-element logical OR operation The element a in the matrix A ij Value represents area r i and r j Whether there are link edges between them, these edges represent the direct interaction and spatial relationship between regions, and the map of the regions is obtained. S13, Hyperedge construction: To capture the high-order relationship between regions, hyperedges are constructed to group multiple regional nodes together. First, based on human mobility, from the perspectives of the source and destination of the region, k2 (k2 = 10) regions closest to each region are selected to form hyperedges. All these hyperedge sets constitute a hyperedge association matrix based on mobility data. Secondly, by calculating the cosine similarity between POI features and selecting the most similar k3 (k3 = 10) nodes to form a hyperedge, the association matrix is ​​constructed similarly Then merge H by concatenation operation f and H p Generate the final hyperedge association matrix H = H f ||H p ,in m=m1+m2; get the hypergraph S14, hybrid graph construction: While completing the street view context embedding and edge construction process, combine the edge set obtained in the previous steps and hyperedge sets Get based on region The mixed graph of is taken as the input of the framework, and the mixed graph is represented as 3. The unsupervised region representation learning method for learning urban region embedding according to claim 2, characterized in that: In step S2, there are three steps including graph / hypergraph enhancement, graph / hypergraph representation and graph / hypergraph contrast loss function, specifically including: S21. Graph / hypergraph enhancement: Use data augmentation techniques to perturb graph and hypergraph structures and generate different views to simulate and alleviate noise problems in graphs and hypergraphs. For graph structures, a random masking matrix is ​​sampled from a Bernoulli distribution with probability p. m Randomly mask some dimensions of the node feature matrix X with probability p r Randomly delete graph Some edges in the original graph Generate enhanced two-graph view and Similarly, for the hypergraph structure, with probability p' m Randomly mask node features in different dimensions with probability p' r Random deletion For some hyperedges in , we get two enhanced hypergraph views and Since nodes in graphs and hypergraphs essentially represent regions, feature masking is the same for both, i.e. S22. Graph / hypergraph representation: Use graph convolutional networks (GCNs) and hypergraph neural networks (HGNNs) as encoders to learn representations of nodes and hyperedges from enhanced graphs and hypergraphs. Graph representation learning focuses on updating the embedding of each node by aggregating neighbor node information. and The input is fed into the shared parameter graph convolutional network GCN encoder to obtain the node embedding matrices Y1 and Y2 of the two views, and then the node embedding matrix is ​​projected into a smaller potential space through a two-layer MLP projection head to obtain the projected embedding matrix and Similarly, hypergraph representation learning is performed on the two generated enhanced views by using a hypergraph neural network (HGNN) encoder to learn node and hyperedge representations. The embeddings of nodes and hyperedges are updated iteratively using the degree matrices of nodes and hyperedges and the trainable weight matrix. Finally, the node embedding matrices P1 and P2 and the hyperedge embedding matrices Q1 and Q2 of the two views are obtained. Two layers of MLP are used as projection heads to project the embeddings of nodes and hyperedges into smaller latent spaces, respectively, to obtain the embedding matrices of nodes and hyperedges of the corresponding views. and S23. Graph / hypergraph contrastive loss: Design a contrastive loss function to guide the model to learn to distinguish the representations of the same region between different views and to bring the representations of similar regions in different views closer. For graph contrastive learning, use the InfoNCE loss function in contrastive learning to calculate the node-level contrastive loss. For any node v i , embed it in the first view As an anchor, the corresponding embedding in the second view As positive samples, other nodes are used as negative samples. The node-level contrast loss is calculated by maximizing the similarity between positive samples (calculated by the cosine similarity function) and minimizing the similarity between the anchor point and the negative samples. Right now in τ v represents the temperature parameter for the penalty of hard negative samples, s(·) is the cosine similarity function, which is used to measure the similarity between two representations. represents the number of nodes; similarly, for hypergraph contrastive learning, node-level, hyperedge-level, and membership-level contrastive losses are designed to capture high-order relationships in the hypergraph and promote learning representations with high-order information; after obtaining the projected node and hyperedge embedding matrices, the node-level, hyperedge-level, and membership-level contrastive losses are calculated respectively; the node-level contrastive loss is calculated by maximizing the number of nodes with the same v i Representation in different views and The similarity of is calculated by minimizing the similarity between different nodes, that is, in The hyperedge-level contrastive loss is obtained by maximizing the same hyperedge e i Representation in different views and The similarity is calculated, that is, in τ e is the temperature parameter, represents the number of hyperedges; while the contrast loss at the affiliation level considers the association between nodes and hyperedges and enhances the representation of nodes and hyperedges, i.e. in τ m is the temperature parameter, Represents node v i The number of hyperedges it belongs to, is the discriminator implemented by a bilinear network; finally, the weighted sum of the three levels of contrast loss is obtained to obtain the total hypergraph contrast loss 4. The unsupervised region representation learning method for learning urban region embedding according to claim 3, characterized in that: In step S3, the quality and expressiveness of region embedding are further optimized by introducing a cross-module contrastive learning mechanism. Specifically, the following steps are involved: S31, information fusion: using a weight-based fusion mechanism to embed nodes from different modules into representation Z v and Z n Perform weighted 2 fusion to obtain the fusion representation Z f =αZ v +(1-α)Z n , that is, respectively and Where α∈(0,1) is a weight parameter used to balance the contribution of different modules; S32. Contrast loss calculation: Introduce fusion contrast loss and use the fused node embedding to represent Z f ,Will and The same node pairs in the region r are regarded as positive pairs, otherwise they are regarded as negative pairs; i As an anchor node, the fusion contrast loss is defined by calculating the contrast loss between the same node pairs Then the contrast losses of all nodes are combined to obtain the overall loss function of cross-module contrastive learning where τ c For l c The corresponding temperature parameters are used to optimize the final node representation.

5. The unsupervised region representation learning method for learning urban region embedding according to claim 4, characterized in that: In step S4, a comprehensive region embedding learning process is implemented to obtain the final region embedding by jointly optimizing the loss functions of all contrastive learning modules; specifically, it includes: The overall loss is defined as Where λ1 and λ2 are the The parameters of importance; the final region representation is obtained after model training In each training cycle, the graph and hypergraph enhancement operations are performed, first from and Generate two enhanced graphs and hypergraphs, which are then fed into the corresponding shared graph encoders f θ (·) and hypergraph encoder to learn relevant representations; subsequently, different projection heads (p φ (·), p ψ (·)and ) maps these embeddings into the latent space and calculates the contrastive loss of the graph and hypergraph modules respectively; next, a controller component is used to fuse the node representations of the graph and hypergraph modules, resulting in a cross-module contrastive loss; finally, the model is trained by optimizing the overall objective loss; after training, the well-trained model will output the region representation E, which is applied to various downstream tasks.

6. An unsupervised region representation learning system for learning urban region embeddings, characterized by: The system adopts the method as claimed in any one of claims 1 to 5.

Images