A multi-view urban area embedding method based on spatial function consistency
By introducing spatial functional consistency graph convolution and cross-view interaction mechanism into multi-view urban region representation learning, the problems of insufficient spatial smoothness and functional consistency in existing methods are solved, achieving more stable and efficient region representation and improving the accuracy and generalization ability of downstream tasks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JILIN UNIVERSITY
- Filing Date
- 2026-06-08
- Publication Date
- 2026-07-03
Smart Images
Figure CN122336073A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of urban computing and data mining technology, specifically to a multi-view urban area embedding method based on spatial functional consistency. Background Technology
[0002] With the rapid growth of urban digitization and multi-source data, urban regional representation learning has gradually become one of the fundamental technologies in smart city applications. Its goal is to learn low-dimensional embeddings using multi-source information from urban areas to support downstream tasks such as land use identification, crime risk prediction, and pedestrian / check-in prediction. Existing methods are typically based on a graph representation learning paradigm: introducing multi-view / multi-source data (such as human travel, POIs, building morphology, road structure, etc.) to construct multi-view graph structures, and then aligning and fusing multi-view representations through strategies such as graph convolution, contrastive learning, or attention fusion to obtain richer regional semantic expressions.
[0003] In multi-view urban region representation learning, common technical approaches include: modeling the relationships between regions within each view (e.g., constructing edge weights based on adjacency matrices or travel flows and performing message passing), and then fusing the region representations from different views (e.g., early fusion, late fusion, or attention fusion); at the same time, there are also methods that use cross-view comparative learning, mutual information maximization, and other objectives to align region representations from different views in a shared space, thereby improving the consistency and robustness of the representations.
[0004] Despite some progress in existing technologies, two major shortcomings remain that severely limit the quality of urban area representations and the accuracy of downstream tasks:
[0005] (1) Single-view modeling relies heavily on feature similarity for composition and propagation. It lacks explicit constraints on spatial proximity priors, making it difficult to balance spatial smoothness and functional consistency. This results in insufficient embedding of adjacent and functionally similar regions, making it more sensitive to noise and sparse observations, and lacking stability and discriminability.
[0006] (2) The interaction scope of cross-view modeling is usually limited to the alignment / fusion of "same region", ignoring the implicit cross-regional associations of "semantically related but different regions", making it difficult to form global semantic connections across regions and views. This results in insufficient cross-view information dissemination, leading to large fluctuations and insufficient transferability in the returns of tasks such as crime prediction, check-in prediction, and land use clustering. Therefore, a multi-view region embedding learning method is needed to simultaneously characterize spatial proximity and functional similarity consistency constraints within the view, and to establish region-region semantic connections and realize global interaction at the cross-view level, so as to obtain stronger task-independent generality and cross-task robustness. Summary of the Invention
[0007] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.
[0008] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:
[0009] S1: Multi-view data collection and processing: Collect multi-view area data of the target city and process the data of each view;
[0010] S2: Spatial Functional Consistency View Representation Learning: Design a spatial functional consistency graph convolution module to fuse spatial distance relationships between regions (spatial proximity) and similarity between regional features (functional similarity) to obtain spatial functional consistency weights. Adaptively reweight the edge relationships in the original graph structure and realize the aggregation and updating of regional information through the graph convolution propagation mechanism.
[0011] S3: Cross-view interaction within the same region: Design a cross-view interaction module within the same region. Take the local representation of the same region obtained in various views as input, model the relationship between different views through attention interaction mechanism, and introduce an adaptive fusion mechanism to dynamically integrate the shared information of the same region in different views.
[0012] S4: Cross-regional cross-view interaction: Design a cross-regional cross-view interaction module to select a set of candidate regions with high semantic relevance from other views for each region, perform cross-regional information aggregation, and merge the cross-regional interaction results with the cross-view interaction results within the region;
[0013] S5: Dual-feature attention fusion: Design a dual-feature attention fusion module, including two stages: view-level fusion and region-level fusion, to obtain unified cross-view fusion features and output the final urban region representation.
[0014] S6: Downstream task application: Obtain the low-dimensional embedding representation of the urban area and use it as a general feature input to different downstream task models.
[0015] As a preferred embodiment of the multi-view urban area embedding method based on spatial functional consistency described in this invention, the multi-view area data in S1 includes pedestrian flow data, point of interest (POI) data, geographic adjacency data, and spatial distance information between areas. A travel view reflecting the flow relationship between areas is constructed based on pedestrian flow data; regional functional features are extracted based on POI category distribution information and a functional view is constructed; and a spatial view is constructed based on geographic adjacency and spatial distance information between areas.
[0016] As a preferred embodiment of the multi-view city region embedding method based on spatial functional consistency described in this invention, the processing method for the multi-view region data is as follows: City region division: dividing the target city into a set of non-overlapping city regions. , in, For the first One region;
[0017] Human travel data: Within a given time window, statistics are collected on travel from a region. To the area The number of travelers, constructing a human travel matrix , where each entry Indicates from arrive The number of travelers;
[0018] Point of interest data: for each region Count the number of different POI categories and construct a POI vector set. ,in For the number of POI categories, Each dimension corresponds to the region Number of POIs in this category;
[0019] Geographic adjacency construction: for any region The set of geographically adjacent regions is represented as ,in for The number of geographical neighbors; the set of all regional geographical neighbors is represented as ;
[0020] Representation learning objective definition: The objective is to learn a low-dimensional embedding representation of a city region, denoted as in For the region of Dimensional embedding representation.
[0021] As a preferred embodiment of the multi-view urban area embedding method based on spatial functional consistency described in this invention, the specific method of S2 is as follows: Within each view, a spatial functional consistency score matrix is calculated. First, spatial proximity and view feature similarity are defined and multiplied to obtain the consistency score, as shown in the following formula:
[0022]
[0023] in, Indicates the area With the region Spatial functional consistency score; Indicates spatial proximity. Indicates the similarity of view features; For the region and Euclidean distance, This is the distance scaling factor; This refers to the feature vector of the region under the corresponding view;
[0024] The aforementioned consistency scores are injected into the adjacency matrix of the GCN to adaptively adjust edge weights, and then normalized. This process is then used to update the GCN layer, resulting in the region representation matrix for each view. ,in Indicates the first The region representation matrix for each view is given by the following formula:
[0025]
[0026]
[0027] in, The adjacency matrix after adding self-loops, The enhanced adjacency matrix after injecting consistency; For learnable scaling parameters; It is a diagonal matrix of degree. This is the normalized adjacency matrix.
[0028] As a preferred embodiment of the multi-view urban area embedding method based on spatial functional consistency described in this invention, the specific method of S3 is to use a multi-head attention mechanism to interact with cross-view representations of the same area, and the calculation formula is as follows:
[0029]
[0030]
[0031] in, The learnable parameter matrix; MultiHeadAttn() is the multi-head attention calculation function, which splits multiple groups of parallel attention heads to extract multi-dimensional features and aggregates them to obtain the multi-head attention output; Attention() represents the calculation of the attention function; These are the query key and value after linear projection, respectively; The scaling factor represents the dimension of the representation; softmax() is the normalization exponential function. This represents the shared representation corresponding to each view;
[0032] By introducing a learnable linear interpolation mechanism, the region is obtained. Shared representation The calculation formula is as follows:
[0033]
[0034] in, For the first Shared representation after the integration of various regions; These are learnable parameters; This represents the original region. This is a representation of the region after attention interaction, and the final output is a set of representations of the region after cross-view interaction. .
[0035] As a preferred embodiment of the multi-view urban area embedding method based on spatial functional consistency described in this invention, the specific method of S4 is as follows: First, construct the cross-view area correlation. Assuming that two areas are highly correlated within their respective views, their cross-view correlation should also be high, the formula is as follows:
[0036]
[0037] in, Indicates the view index and The feature similarity matrix within view p. Represents a view Central region With View Central region Cross-view, cross-region correlation;
[0038] For each view area in From another view Select the view with the highest cross-view relevance A set of candidate regions And calculate the scaled dot product attention weights on this set, as shown in the following formula;
[0039]
[0040] in, The total number of regions is represented by , and TopK represents the number of regions before filtering. A method with one parameter; Indicates the inner product; The scaling factor is used to prevent the inner product value from being too large when the feature dimension is too high, which would cause exp() to fall into the saturation range and the gradient to vanish, and to stabilize the attention weight distribution after Softmax normalization. The normalized attention weights are obtained by weighted aggregation of semantically relevant region information from other views, as shown in the following formula:
[0041]
[0042] By fusing cross-region and cross-view aggregation results with intra-region view interaction results, a globally interactive region representation is obtained. The formula is as follows:
[0043]
[0044] in, For view Central region Cross-region aggregation representation; The learnable fusion coefficients are given; the output is... .
[0045] As a preferred embodiment of the multi-view urban area embedding method based on spatial functional consistency described in this invention, in step S5, view fusion involves calculating arbitrary view pairs. In the region Fusion score And obtain the view weight. Cross-view fusion representation The formula is as follows:
[0046]
[0047]
[0048] in, For the region In view The interaction indicates; is a linear transformation matrix; "||" indicates concatenation; LeakyReLU() is a linear activation function with leakage correction, which introduces a minimum fixed slope in the negative input interval to solve the problem of neuron death caused by the negative half-axis gradient returning to zero in traditional ReLU, and realizes non-linear mapping of features; To normalize view weights; The fusion representation matrix is used; softmax() is a normalized exponential function that outputs the probability distribution or attention weights.
[0049] Region Fusion: To further model the interaction relationships between regions based on the fused features, this invention employs a self-attention mechanism with residual connections to obtain the final fused region representation, which can be stacked. Each block is used to achieve deep feature interaction; ultimately, a set of region embeddings is obtained. The formula is as follows:
[0050]
[0051]
[0052] Among them, LayerNorm() is the layer normalization function, which standardizes the feature dimension of a single sample to accelerate the convergence of the neural network and stabilize the training distribution; softmax() is the normalization exponential function, which outputs the probability distribution or attention weights; MLP() is a multilayer perceptron, which is a basic neural network composed of fully connected layers, used to perform nonlinear transformations and dimension mappings on features; Dropout() is a random deactivation function, which randomly discards some neurons during the training of the neural network with a set probability to prevent the model from overfitting and improve generalization ability; E represents the finally learned urban area embedding representation.
[0053] As a preferred embodiment of the multi-view urban area embedding method based on spatial functional consistency described in this invention, step S5 further includes a self-supervised learning objective and optimization, the specific method of which is as follows:
[0054] Decoding: The learned region representation Decoded via view-specific decoder , where superscript These correspond to the geographic neighbor view, trip source view, trip destination view, and POI view, respectively.
[0055] Geographic Neighbor Loss: Based on Positive / Negative Geographic Neighbor Sample Definition The formula is as follows:
[0056]
[0057] in, For the region The final embedding; and From the region The positive geographical neighbors and negative samples are represented by the samples. It is a norm 2;
[0058] Trip Restructuring Loss: First calculate the original trip distribution Represented as Departure Arrival The prior probabilities are then used to reconstruct the source distributions. With target distribution The formula is as follows:
[0059]
[0060]
[0061]
[0062]
[0063] in, Given the decoded source / target representation of the trip, exp(.) is an exponential function, yielding... , middle and , respectively, represent the real and reconstructed distributions;
[0064] POI reconstruction loss: given the POI similarity matrix Represented by POI ,definition ,in, For the region and POI similarity; To reconstruct the product, the formula is as follows:
[0065]
[0066] Overall objective: To obtain the overall optimization objective by summing the three types of losses. ,in, Constrain spatial neighbor structure consistency Constraints on the reconfigurability of travel distribution The reconfigurability of the Points of Interest (POI) correlation, with the overall loss function as follows:
[0067] .
[0068] As a preferred embodiment of the multi-view urban area embedding method based on spatial functional consistency described in this invention, the downstream task model in S6 is used for crime prediction, pedestrian flow prediction, land use reasoning, and attendance prediction.
[0069] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention can collaboratively constrain the spatial proximity and functional similarity between regions within the view, and adaptively enhance and reweight the edge relationships in the original graph structure, thereby effectively suppressing information interference caused by noisy edges and weakly correlated connections, and improving the accuracy and stability of local structure modeling of regions; at the same time, this invention breaks through the technical path of existing technologies where cross-view interaction is mainly limited to the alignment of the same region, and further establishes semantic association relationships between different views and different regions, enhancing the global semantic connectivity and cross-view information propagation capabilities of region representation; on this basis, through a two-layer attention fusion mechanism, adaptive integration of multi-view features and inter-region association features is achieved, further improving the discriminativeness, robustness and generalization ability of the final region embedding representation.
[0070] Experimental results show that this invention outperforms existing baseline methods in various downstream tasks, including crime prediction, land use clustering, and attendance prediction. Specifically, in crime prediction, the mean absolute error is reduced by 14.09%, and the root mean square error is reduced by 7.02%; in land use clustering, the adjusted RAND index is improved by 15.90%; and in attendance prediction, the mean absolute error is reduced by 18.83%, and the root mean square error is reduced by 15.49%. Furthermore, ablation experiments further validate the necessity of each core technology module for improving overall performance, and training trend analysis results indicate that this invention also exhibits faster convergence speed and higher training stability. Attached Figure Description
[0071] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:
[0072] Figure 1 This is a schematic diagram of the overall process of the Urban Region Representation Learning Framework (SFC-CVRI) proposed in this invention;
[0073] Figure 2 This is a schematic diagram of the core model structure of the SFC-CVRI framework proposed in this invention;
[0074] Figure 3 The diagram shows the impact of four key hyperparameters of this invention on model performance. (a) shows the impact of the top-k number of cross-view interaction regions on model performance; (b) shows the spatial decay parameter. The impact on model performance is shown in the diagram; (c) is the spatial functional consistency scaling factor. (d) shows the impact of the number of SFC-GCN layers on model performance. Detailed Implementation
[0075] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0076] Secondly, the present invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of the present invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not according to the usual scale. Furthermore, the schematic diagrams are merely examples and should not limit the scope of protection of the present invention. In addition, actual fabrication should include three-dimensional spatial dimensions of length, width, and depth.
[0077] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0078] Please see Figure 1 The figure illustrates the overall implementation process of the present invention, including urban area data collection, data storage, data cleaning, feature preprocessing, dataset construction, SFC-CVRI model construction and training, model performance evaluation, pre-trained model generation, and the generation of regional representations using the pre-trained model and their application to downstream tasks. It demonstrates the complete technical chain of the present invention from raw data input to model training and then to model application output.
[0079] Specifically, this invention proposes an urban region representation learning framework based on spatial-functional consistency and cross-view region interaction, called SFC-CVRI (Spatial-Functional Consistency and Cross-View RegionInteraction for Urban Region Representation Learning). It includes the following steps:
[0080] S1: Multi-view Data Collection and Processing: Collect multi-view regional data for the target city, including but not limited to pedestrian flow data, Points of Interest (POI) data, geographic adjacency data, and spatial distance information between regions. First, the target city is divided into regions, and a unified urban regional index system is constructed to ensure consistency in regional granularity across different data sources. Subsequently, each view's data undergoes cleaning, noise reduction, missing value handling, standardization, and feature preprocessing. Specifically: a travel view reflecting inter-regional flow relationships is constructed based on pedestrian flow data; regional functional features are extracted based on POI category distribution information, and a functional view is constructed; spatial views are constructed based on inter-regional geographic adjacency and spatial distance information. Finally, a regional feature matrix and intra-view relationship graph are generated for each view, providing a unified data foundation for subsequent model training and inference.
[0081] S2: Spatial-Functional Consistency In-View Representation Learning: Addressing the challenge of existing methods simultaneously characterizing spatial proximity and functional similarity relationships between regions within a single view, this invention designs a spatial-functional consistency learning module, namely the Spatial-Functional Consistency Graph Convolution Network (SFC-GCN). Within each view, this module characterizes spatial proximity based on spatial distance relationships between regions and functional similarity based on the similarity between region features, fusing these two aspects to obtain spatial-functional consistency weights. Based on this, the edge relationships in the original graph structure are adaptively reweighted using these consistency weights, and then the aggregation and updating of region information are achieved through a graph convolution propagation mechanism. This step allows spatially close regions with high functional correlation to be closer in the representation space, effectively suppressing interference from noisy edges and weakly correlated connections, resulting in more robust in-view representations of local regions.
[0082] S3: Cross-view Interaction within the Same Region: To achieve information sharing and semantic complementarity of the same region across different views, this invention designs a cross-view interaction module for the same region. This module takes the local representations of the same region obtained in various views as input, and models the correlation between different views through an attention interaction mechanism, enabling each view to selectively absorb effective information from other views, thereby enhancing the consistency and completeness of the multi-view representation of the same region. Furthermore, to balance the contributions between the original view representation and the cross-view interaction representation, this invention introduces an adaptive fusion mechanism to dynamically integrate the shared information of the same region in different views, outputting an enhanced representation after cross-view interaction within the region. This step effectively alleviates the impact of insufficient information in a single view or local noise, improving the robustness of the semantic representation of the region.
[0083] S4: Cross-View Interaction Between Regions: To overcome the limitations of existing cross-view interactions, which are typically confined to information exchange within the same region and struggle to uncover potential cross-view semantic relationships between different regions, this invention further designs a cross-regional cross-view interaction module. This module, across different views, selects a set of candidate regions with high semantic relevance for each region from other views, and then performs cross-regional information aggregation. Specifically, by constructing cross-view region correlations, it identifies regions with potential functional or structural connections across different views, and uses an attention aggregation mechanism to weighted summarize the relevant region information, thereby achieving cross-view semantic interaction between different regions. Subsequently, the cross-regional interaction results are fused with the intra-regional cross-view interaction results to obtain a region representation containing richer global semantic relationships. Through this step, the model's learning scope can be expanded from local multi-view within a single region. Figure 1The consistency is extended to global cross-view semantic associations across regions, enhancing the ability of urban area representation to depict complex spatial functional structures.
[0084] S5: Dual-Attention Fusion: To further integrate multi-view interaction results and form a unified, high-quality region representation, this invention designs a dual-attention fusion module (DAFusion). This module includes two stages: view-level fusion and region-level fusion. First, in the view-level fusion stage, the region representations after interaction from each view are weighted and integrated to adaptively learn the contribution of different views to the final region representation, obtaining unified cross-view fusion features. Subsequently, in the region-level fusion stage, the deep relationships between regions are further modeled, and the fused region features are globally interacted and enhanced to output the final city region representation. Through the above two-layer fusion mechanism, the information complementarity between different views and the overall correlation between regions can be taken into account simultaneously, thereby improving the discriminative and generalization capabilities of the region embedding representation.
[0085] S6: Downstream Task Application: After completing the above representation learning, a low-dimensional embedding representation of the urban area is obtained. This regional representation can be used as a general feature input to different downstream task models to support various urban computing tasks, including but not limited to crime prediction, pedestrian flow prediction, land use inference, and attendance prediction. For different task requirements, the regional representation can be directly input into prediction models, classification models, or clustering models for training and inference, thereby verifying the effectiveness and transferability of the learned regional representation in various practical application scenarios.
[0086] Please see Figure 2 This diagram illustrates the hierarchical relationships and data flow among the modules within the model. It includes using pedestrian traffic data, POI data, and geographic neighbor data as multi-view inputs, completing intra-view representation learning via the Spatial Functional Consistency Learning Module (SFC-GCN), then achieving cross-view semantic information interaction through intra-regional cross-view interaction and cross-view / cross-regional interaction modules. Subsequently, a dual-feature attention fusion module completes view-level fusion and region-level fusion, ultimately outputting a general urban regional representation to serve downstream tasks. This diagram reflects the core technical structure of this invention across three levels: "intra-view consistency modeling—cross-view interaction—fusion output."
[0087] The specific model training includes the following steps:
[0088] S1: Multi-view data collection and processing:
[0089] Urban area division: Dividing the target city into a set of non-overlapping urban areas. , in, For the first Each region.
[0090] Human travel data: Within a given time window, statistics are collected on travel from a region. To the area The number of travelers, constructing a human travel matrix , where each entry Indicates from arrive The number of travelers.
[0091] POI data: for each region Count the number of different POI categories and construct a POI vector set. ,in For the number of POI categories, Each dimension corresponds to the region The number of POIs in this category.
[0092] Geographic adjacency construction: for any region The set of geographically adjacent regions is represented as ,in for The number of geographical neighbors; the set of all regional geographical neighbors is represented as .
[0093] Representation learning objective definition: The objective of this invention is to learn a low-dimensional embedding representation of an urban region, denoted as in For the region of Dimensional embedding representation.
[0094] S2: In-view spatial functional consistency modeling:
[0095] (1) Within each view, calculate the spatial functional consistency score matrix. First, define the spatial proximity and view feature similarity and multiply them to obtain the consistency score, as shown in the following formula:
[0096]
[0097] in, Indicates the area With the region Spatial functional consistency score; Indicates spatial proximity. Indicates the similarity of view features; For the region and Euclidean distance, This is the distance scaling factor; This is the feature vector of the region in the corresponding view.
[0098] (2) The above consistency scores are injected into the adjacency matrix of the GCN to adaptively adjust the edge weights, and normalized. The GCN layer is then updated to obtain the region representation matrix of each view. ,in Indicates the first The region representation matrix for each view is given by the following formula:
[0099]
[0100]
[0101] in, The adjacency matrix after adding self-loops, The enhanced adjacency matrix after injecting consistency; For learnable scaling parameters; It is a diagonal matrix of degree. This is the normalized adjacency matrix.
[0102] S3: Cross-view interaction within the region:
[0103] (1) To achieve information sharing of the same region across different views, this invention employs a multi-head attention mechanism to interact with the cross-view representation of the same region. The calculation formula is as follows:
[0104]
[0105]
[0106] in, The learnable parameter matrix; MultiHeadAttn() is the multi-head attention calculation function, which splits multiple groups of parallel attention heads to extract multi-dimensional features and aggregates them to obtain the multi-head attention output; Attention() represents the calculation of the attention function; These are the query key and value after linear projection, respectively; The scaling factor represents the dimension of the representation; softmax() is the normalization function. This represents the shared representation corresponding to each view.
[0107] (2) To achieve an adaptive balance between the original view representation and the attention-fused representation, this invention introduces a learnable linear interpolation mechanism to obtain the region. Shared representation The calculation formula is as follows:
[0108]
[0109] in, For the first Shared representation after the integration of various regions; These are learnable parameters; This represents the original region. This is a representation of the region after attention interaction, and the final output is a set of representations of the region after cross-view interaction. .
[0110] S4: Cross-view interaction between regions:
[0111] (1) To characterize the potential semantic relationships between different regions under different views, this invention first constructs cross-view region correlation. It is assumed that if two regions are highly correlated within their respective views, their cross-view correlation should also be high. The formula is as follows:
[0112]
[0113] in, Indicates the view index and The feature similarity matrix within view p. Represents a view Central region With View Central region Cross-view and cross-region correlation.
[0114] (2) For each view area in From another view Select the view with the highest cross-view relevance A set of candidate regions And calculate the scaled dot product attention weights on this set, as shown in the following formula;
[0115]
[0116] in, The total number of regions is represented by , and TopK represents the number of regions before filtering. A method with one parameter; Indicates the inner product; The scaling factor is used to prevent the inner product value from being too large when the feature dimension is too high, which would cause exp() to fall into the saturation range and the gradient to vanish, and to stabilize the attention weight distribution after Softmax normalization. The attention weights are normalized. They are obtained by weighted aggregation of semantically relevant region information from other views, as shown in the following formula:
[0117]
[0118] (3) Merge the cross-region and cross-view aggregation results with the intra-region view interaction results to obtain the global interactive region representation. The formula is as follows:
[0119]
[0120] in, For view Central region Cross-region aggregation representation; The learnable fusion coefficients are given; the output is... .
[0121] S5: Dual-feature attention fusion:
[0122] (1) View fusion: Calculate any view pair In the region Fusion score And obtain the view weight. Cross-view fusion representation The formula is as follows:
[0123]
[0124]
[0125] in, For the region In view The interaction indicates; This is a linear transformation matrix; "||" indicates concatenation. To normalize view weights; The fusion representation matrix is used; LeakyReLU() is a linear activation function with leakage correction, which introduces a very small fixed slope in the negative input interval to solve the problem of neuron death caused by the gradient of the negative half axis of traditional ReLU returning to zero, and realizes non-linear mapping of features; softmax() is a normalized exponential function that outputs a probability distribution or attention weights.
[0126] (2) Region fusion: To further model the interaction relationships between regions based on the fused features, this invention adopts a self-attention mechanism with residual connections to obtain the final fused region representation, which can be stacked. Each block is used to achieve deep feature interaction; ultimately, a set of region embeddings is obtained. The formula is as follows:
[0127]
[0128]
[0129] Here, LayerNorm() is the layer normalization function, which standardizes the dimension of single-sample features to accelerate neural network convergence and stabilize the training distribution; softmax() is the normalization exponential function, which outputs the probability distribution or attention weights; MLP() is a multilayer perceptron, a basic neural network composed of fully connected layers, used to perform nonlinear transformations and dimension mappings on features; Dropout() is a random deactivation function, which randomly discards some neurons during neural network training with a set probability to prevent overfitting and improve generalization ability. E represents the finally learned urban region embedding representation.
[0130] S6: Self-supervised learning objectives and optimization:
[0131] (1) Decoding: Representing the learned regions Decoded via view-specific decoder Among them, superscript These correspond to the geographic neighbor view, trip source view, trip destination view, and POI view, respectively.
[0132] (2) Geographic neighbor loss: based on the definition of positive / negative geographic neighbor samples The formula is as follows:
[0133]
[0134] in, For the region The final embedding; and From the region The positive geographical neighbors and negative samples are represented by the samples. It is a 2-norm.
[0135] (3) Trip restructuring loss: First calculate the original trip distribution Represented as Departure Arrival The prior probabilities are then used to reconstruct the source distributions. With target distribution The formula is as follows:
[0136]
[0137]
[0138]
[0139]
[0140] in, Given the decoded source / target representation of the trip, exp(.) is an exponential function, yielding... , middle and , which represent the true and reconstructed distributions, respectively.
[0141] (4) POI reconstruction loss: given the POI similarity matrix Represented by POI ,definition .in, For the region and POI similarity; To reconstruct the product, the formula is as follows:
[0142]
[0143] (5) Overall objective: The overall optimization objective is obtained by adding the three types of losses together. ,in, Constrain spatial neighbor structure consistency Constraints on the reconfigurability of travel distribution The reconfigurability of the Points of Interest (POI) correlation, with the overall loss function as follows:
[0144]
[0145] S7: Downstream Task Applications of the Model
[0146] Applications of prediction and clustering tasks: embedding regions The features are input to downstream predictors (such as linear regression / MLP / temporal series predictors) for training and inference, used for tasks such as crime intensity prediction, attendance prediction, and land use type clustering. The downstream predictor is denoted as... Its input consists of region embeddings and other data.
[0147] Experimental results
[0148] Table 1: Overall performance comparison across three downstream tasks
[0149]
[0150] As shown in Table 1, the experimental results demonstrate the significant technical effectiveness of the proposed framework. Validation was conducted on the Manhattan dataset, focusing on three downstream tasks: crime prediction, land use clustering, and check-in prediction. For the regression task, MAE, RMSE, and R² were used as evaluation metrics, while for the clustering task, NMI and ARI were used. Overall, the proposed method outperforms various single-view and multi-view baseline methods in all three tasks. In the crime prediction task, the model achieved MAE=56.889, RMSE=78.972, and R²=0.715, representing a 14.09% decrease in MAE and a 7.02% decrease in RMSE compared to the strongest baseline. In the land use clustering task, the model achieved NMI=0.816 and ARI=0.672, with an ARI improvement of 15.90%. In the check-in prediction task, the model achieved MAE=219.283, RMSE=347.512, and R²=0.820, representing an 18.83% decrease in MAE and a 15.49% decrease in RMSE compared to the baseline. These results demonstrate that the proposed SFC-CVRI framework can consistently achieve superior performance across different types of downstream tasks, exhibiting strong task generalization ability and representation versatility.
[0151] Table 2: Experimental Results of Modular Ablation
[0152]
[0153] Furthermore, ablation experiments verified the necessity of each component module. As shown in Table 2, after removing the SFC-GCN, the R² for crime prediction decreased from 0.715 to 0.638, and the R² for check-in prediction decreased from 0.820 to 0.743, indicating that spatial function consistency modeling plays a crucial role in maintaining stable intra-view structure and improving prediction performance. After removing the cross-view interaction module within the region, the ARI for land use clustering decreased significantly from 0.672 to 0.520, indicating that cross-view semantic alignment within the same region is crucial for improving clustering quality. After removing the cross-view interaction module between regions, multiple task indicators decreased, indicating that cross-regional and cross-view semantic association helps enhance global semantic connectivity. However, if the dual-feature attention fusion is replaced with simple additive fusion, all three tasks show significant degradation, indicating that an effective view fusion mechanism is also an important guarantee for obtaining high-quality regional representations.
[0154] like Figure 3 As shown, top-k represents the number of semantically relevant regions selected by each region from other views during cross-view and cross-region interaction. A small top-k may lead to insufficient interactive information, making it difficult to fully capture cross-view semantic relationships; while an excessively large top-k may introduce weakly relevant or even noisy regions, thus affecting the representation learning effect. Used to control the decay rate of spatial distance in spatial functional consistency modeling. When When the size is small, the spatial influence range is narrow, and the model focuses more on areas with very close proximity; when When the value is large, distant regions will also receive higher spatial correlation weights, which may lead to over-smoothing. It is an adjustment coefficient when the spatial function consistency score is injected into the graph convolutional adjacency matrix, used to control the strength of the influence of spatial function prior information on graph message passing. Too small a space will weaken the consistency of its functions. An excessively large number of layers may cause the model to rely too heavily on prior relationships, reducing the flexibility of representation learning. The number of layers in SFC-GCN represents the depth of graph convolution propagation. Too few layers make it difficult to capture high-order neighborhood information, while too many layers may cause oversmoothing problems, making the representations of different regions tend to be similar.
[0155] Regarding evaluation metrics, this paper employs different performance metrics for different downstream tasks. For crime prediction and attendance prediction tasks, the coefficient of determination (R²) is primarily used as the evaluation metric. R² measures the explanatory power of the model's predicted values for changes in actual values; a higher R² indicates better prediction performance, and a closer R² to 1 indicates a stronger fit to the number of crimes or attendances in the region. For land use clustering tasks, the adjusted RAND index (ARI) is used as the main evaluation metric. ARI measures the consistency between the clustering results and the actual categories, and corrects for random clustering results; a higher ARI indicates that the clustering results are closer to the actual land use types. Therefore, in the hyperparameter sensitivity experiment, R² is used to reflect regression performance for crime prediction and attendance prediction, while ARI is used to reflect clustering quality for land use clustering. This allows for a comprehensive evaluation of the model's stability and generalization ability under different hyperparameter settings from different task perspectives.
[0156] In summary, the SFC-CVRI framework proposed in this invention revolves around a complete technical chain: "inputting multi-view urban area data—modeling spatial function consistency within views—cross-view interaction within regions—cross-view interaction between regions—dual-feature attention fusion—outputting a unified urban area representation and serving downstream tasks." Addressing the issues of unstable local structures and insufficient global cross-view semantic connections in existing multi-view area representation learning, it specifically solves these problems through modules on spatial function consistency, cross-view interaction within regions, cross-view interaction between regions, and dual-feature attention. Ultimately, it enhances the spatial function consistency of area representations, improves global semantic connectivity, and enhances cross-task generalization ability. Experimental results fully verify the effectiveness, robustness, and practical application value of this invention in real-world urban multi-task scenarios.
[0157] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.
Claims
1. A multi-view urban area embedding method based on spatial functional consistency, characterized in that, Includes the following steps: S1: Multi-view data collection and processing: Collect multi-view area data of the target city and process the data of each view; S2: Spatial Functional Consistency View Representation Learning: Design a spatial functional consistency graph convolution module to fuse spatial distance relationships between regions (spatial proximity) and similarity between regional features (functional similarity) to obtain spatial functional consistency weights. Adaptively reweight the edge relationships in the original graph structure and realize the aggregation and updating of regional information through the graph convolution propagation mechanism. S3: Cross-view interaction within the same region: Design a cross-view interaction module within the same region. Take the local representation of the same region obtained in various views as input, model the relationship between different views through attention interaction mechanism, and introduce an adaptive fusion mechanism to dynamically integrate the shared information of the same region in different views. S4: Cross-regional cross-view interaction: Design a cross-regional cross-view interaction module to select a set of candidate regions with high semantic relevance from other views for each region, perform cross-regional information aggregation, and merge the cross-regional interaction results with the cross-view interaction results within the region; S5: Dual-feature attention fusion: Design a dual-feature attention fusion module, including two stages: view-level fusion and region-level fusion, to obtain unified cross-view fusion features and output the final urban region representation. S6: Downstream task application: Obtain the low-dimensional embedding representation of the urban area and use it as a general feature input to different downstream task models.
2. The multi-view urban area embedding method based on spatial functional consistency according to claim 1, characterized in that, The multi-view area data in S1 includes pedestrian flow data, point of interest data, geographic adjacency data, and spatial distance information between areas. A travel view reflecting the flow relationship between areas is constructed based on pedestrian flow data. Extract regional functional features based on POI category distribution information and construct a functional view; A spatial view is constructed based on the geographical adjacency and spatial distance information between regions.
3. The multi-view urban area embedding method based on spatial functional consistency according to claim 2, characterized in that, The processing method for the multi-view area data is as follows: City area division: Divide the target city into a set of non-overlapping city areas. , in, For the first One region; Human travel data: Within a given time window, statistics are collected on travel from a region. To the area The number of travelers, constructing a human travel matrix , where each entry Indicates from arrive The number of travelers; Point of interest data: for each region Count the number of different POI categories and construct a POI vector set. ,in For the number of POI categories, Each dimension corresponds to the region Number of POIs in this category; Geographic adjacency construction: for any region The set of geographically adjacent regions is represented as ,in for The number of geographical neighbors; the set of all regional geographical neighbors is represented as ; Representation learning objective definition: The objective is to learn a low-dimensional embedding representation of a city region, denoted as in For the region of Dimensional embedding representation.
4. The multi-view urban area embedding method based on spatial functional consistency according to claim 1, characterized in that, The specific method of S2 is as follows: Within each view, calculate the spatial functional consistency score matrix. First, define spatial proximity and view feature similarity and multiply them to obtain the consistency score, as shown in the following formula: in, Indicates the area With the region Spatial functional consistency score; Indicates spatial proximity. Indicates the similarity of view features; For the region and Euclidean distance, This is the distance scaling factor; This refers to the feature vector of the region under the corresponding view; The aforementioned consistency scores are injected into the adjacency matrix of the GCN to adaptively adjust edge weights, and then normalized. This process is then used to update the GCN layer, resulting in the region representation matrix for each view. ,in Indicates the first The region representation matrix for each view is given by the following formula: in, The adjacency matrix after adding self-loops, The enhanced adjacency matrix after injecting consistency; For learnable scaling parameters; It is a diagonal matrix of degree. This is the normalized adjacency matrix.
5. The multi-view urban area embedding method based on spatial functional consistency according to claim 1, characterized in that, The specific method of S3 is to use a multi-head attention mechanism to interact with cross-view representations of the same region, and the calculation formula is as follows: in, The learnable parameter matrix; MultiHeadAttn() is the multi-head attention calculation function, which splits multiple groups of parallel attention heads to extract multi-dimensional features and aggregates them to obtain the multi-head attention output; Attention() represents the calculation of the attention function; These are the query, key, and value after linear projection, respectively; The scaling factor represents the dimension of the representation; softmax() is the normalization function. This represents the shared representation corresponding to each view; By introducing a learnable linear interpolation mechanism, the region is obtained. Shared representation The calculation formula is as follows: in, For the first Shared representation after the integration of various regions; These are learnable parameters; This represents the original region. This is a representation of the region after attention interaction, and the final output is a set of representations of the region after cross-view interaction. .
6. The multi-view urban area embedding method based on spatial functional consistency according to claim 1, characterized in that, The specific method of S4 is as follows: First, construct the cross-view region correlation. It is assumed that if two regions are highly correlated within their respective views, their cross-view correlation should also be high. The formula is as follows: in, Indicates the view index and The feature similarity matrix within view p. Represents a view Central region With View Central region Cross-view, cross-region correlation; For each view area in From another view Select the view with the highest cross-view relevance A set of candidate regions And calculate the scaled dot product attention weights on this set, as shown in the following formula; in, The total number of regions is represented by , and TopK represents the number of regions before filtering. A method with one parameter; Indicates the inner product; The scaling factor is used to prevent the inner product value from being too large when the feature dimension is too high, which would cause exp() to fall into the saturation range and the gradient to vanish, and to stabilize the attention weight distribution after Softmax normalization. The normalized attention weights are obtained by weighted aggregation of semantically relevant region information from other views, as shown in the following formula: By fusing cross-region and cross-view aggregation results with intra-region view interaction results, a globally interactive region representation is obtained. The formula is as follows: in, For view Central region Cross-region aggregation representation; The learnable fusion coefficients are given; the output is... .
7. The multi-view urban area embedding method based on spatial functional consistency according to claim 1, characterized in that, In S5, view fusion: calculates any view pair In the region Fusion score And obtain the view weight. Cross-view fusion representation The formula is as follows: in, For the region In view The interaction indicates; is a linear transformation matrix; "||" indicates concatenation; LeakyReLU() is a linear activation function with leakage correction, which introduces a very small fixed slope in the negative input interval to solve the problem of neuron death caused by the gradient of the negative half axis of traditional ReLU returning to zero, and realizes non-linear mapping of features; To normalize view weights; The fusion representation matrix is used; softmax() is a normalized exponential function that outputs the probability distribution or attention weights. Region Fusion: To further model the interaction relationships between regions based on the fused features, this invention employs a self-attention mechanism with residual connections to obtain the final fused region representation, which can be stacked. Each block is used to achieve deep feature interaction; ultimately, a set of region embeddings is obtained. The formula is as follows: Among them, LayerNorm() is the layer normalization function, which standardizes the feature dimension of a single sample to accelerate the convergence of the neural network and stabilize the training distribution; softmax() is the normalization exponential function, which outputs the probability distribution or attention weights; MLP() is a multilayer perceptron, which is a basic neural network composed of fully connected layers, used to perform nonlinear transformations and dimension mappings on features; Dropout() is a random deactivation function, which randomly discards some neurons during the training of the neural network with a set probability to prevent the model from overfitting and improve generalization ability; E represents the finally learned urban area embedding representation.
8. The multi-view urban area embedding method based on spatial functional consistency according to claim 7, characterized in that, S5 also includes a self-supervised learning objective and optimization, the specific methods of which are as follows: Decoding: The learned region representation Decoded via view-specific decoder , where superscript These correspond to the geographic neighbor view, trip source view, trip destination view, and POI view, respectively. Geographic Neighbor Loss: Based on the definition of positive / negative geographic neighbor samples The formula is as follows: in, For the region The final embedding; and From the region The positive geographical neighbors and negative samples are represented by the samples. It is a 2-norm; Trip Restructuring Loss: First calculate the original trip distribution Represented as Departure Arrival The prior probabilities are then used to reconstruct the source distributions. With target distribution The formula is as follows: in, Given the decoded source / target representation of the trip, exp(.) is an exponential function, yielding... , middle and , respectively, represent the real and reconstructed distributions; POI reconstruction loss: given the POI similarity matrix Represented by POI ,definition ,in, For the region and POI similarity; To reconstruct the product, the formula is as follows: Overall objective: To obtain the overall optimization objective by summing the three types of losses. ,in, Constrain spatial neighbor structure consistency Constraints on the reconfigurability of travel distribution The reconfigurability of the Points of Interest (POI) correlation, with the overall loss function as follows: 。 9. The multi-view urban area embedding method based on spatial functional consistency according to claim 1, characterized in that, The S6 downstream task model is used for crime prediction, pedestrian flow prediction, land use reasoning, and attendance prediction.