Hybrid graph convolutional neural network traffic cell modeling method and system with attention mechanism fusion

Abstract

This invention discloses a hybrid graph convolutional neural network (GCN) traffic cell modeling method and system that incorporates an attention mechanism. The method includes steps such as graph data definition and preprocessing, node feature mapping, neighborhood information aggregation, graph convolutional propagation, network layer stacking, and model training. In the neighborhood information aggregation stage, a single-head attention mechanism is introduced to differentially weight different neighbor nodes. This improves the ability to express potential travel attraction relationships and spatial heterogeneity between traffic cells while maintaining the stability of the GCN structure's normalized propagation. This invention significantly enhances the spatial feature expression and interpretation capabilities while maintaining model stability and training efficiency. It can be widely applied to travel destination prediction, urban functional zoning, traffic attraction analysis, and intelligent transportation system construction, providing technical support and decision-making basis for traffic planning, travel demand analysis, and public transportation optimization.

Description

Technical Field

[0001] This invention relates to the field of traffic zone modeling and analysis technology, specifically to a method and system for traffic zone modeling using a hybrid graph convolutional neural network that incorporates an attention mechanism. Background Technology

[0002] With the continuous advancement of urbanization and the development of traffic sensing facilities, mobile internet, and multi-source traffic big data, urban mobility systems exhibit significant characteristics such as "multi-scale spatial coupling, strong heterogeneity, and dynamic changes." When conducting tasks such as travel demand forecasting, destination attraction pattern analysis, regional function identification, and congestion management, transportation planning and management departments often need to model the interaction relationships between urban traffic zones (i.e., dividing urban space into several relatively homogeneous regional units). The connections between traffic zones depend not only on spatial adjacency but are also influenced by multiple factors, including road accessibility, land use complementarity, public transportation supply, travel costs, and historical interaction intensity, exhibiting significant nonlinearity and heterogeneity. Traditional statistical models (such as multinomial Logit and nested Logit models) have a strong theoretical explanatory basis, but they have limitations in modeling high-dimensional feature interactions, complex nonlinear relationships, and large-scale regional networks, making it difficult to fully characterize the multi-dimensional coupling patterns in real-world transportation systems.

[0003] Deep Neural Networks (DNNs) typically treat samples as independent inputs, making it difficult to explicitly utilize the topological structure and spatial dependencies between traffic zones. Furthermore, their internal decision-making processes often lack clear mechanistic explanations, hindering the formation of verifiable and reusable decision-making bases, thus limiting their widespread application in traffic planning and management scenarios. Graph Convolutional Networks (GCNs) typically use fixed normalized weights determined by adjacency and degree matrices during the information aggregation stage, assuming that neighboring nodes contribute equally to the central node or that the contribution is solely determined by degree. This fails to reflect the differences in functional attributes, travel attractiveness, and interaction intensity between traffic zones. In reality, even if two traffic zones are geographically adjacent, their functional types (e.g., residential, commercial, educational, industrial) and travel interaction intensity may differ significantly. Using an averaging aggregation mechanism can easily dilute important neighbor information, thus affecting the accuracy of characterizing heterogeneous spatial relationships and the effectiveness of downstream tasks. Furthermore, static adjacency relationships may fail to reflect traffic-specific phenomena such as weak cross-community connections, asymmetric attraction, or dynamic accessibility changes, thus limiting the model's ability to express complex spatial behaviors. In traffic area graphs with large node scales and sparse edge relationships, the computational cost of attention networks (GAT) increases significantly, resulting in higher engineering deployment and parameter tuning costs. Compared to the structure-normalized propagation rule of GCN, the pure attention propagation method affects training stability and generalization performance, especially in cases of high noise or insufficient samples, which makes it more prone to overfitting and unsuitable for efficient application in large-scale traffic area graphs. Summary of the Invention

[0004] To address the shortcomings of the prior art, this invention provides a hybrid graph convolutional neural network traffic cell modeling method and system that integrates an attention mechanism. It establishes a hybrid model—Hybrid Attention Graph Convolutional Network (HA-GCN)—that integrates an attention mechanism and graph convolutional propagation. This introduces a lightweight, single-head attention mechanism to differentially weight neighbor nodes, enhancing the ability to express heterogeneous travel attraction relationships between traffic cells while maintaining structural propagation stability. This also provides support for subsequent interpretable analysis and engineering deployment.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A hybrid graph convolutional neural network method for traffic cell modeling incorporating an attention mechanism includes the following steps:

[0007] S1. Obtain traffic zone data for the target city and construct a graph structure with traffic zones as nodes and spatial adjacency or travel reachability relationships as edges. For each traffic zone node... Construct the corresponding original feature vectors and perform preprocessing to form a vector containing traffic zone nodes. Graph data including adjacency or travel relationships, and traffic community node attribute characteristics;

[0008] S2, For each traffic zone node The original feature vectors are used for feature construction and linear mapping. Through linear transformation, the attribute features of traffic cell nodes are mapped to an embedding space of uniform dimension, thus obtaining the traffic cell node... The initial embedding vector;

[0009] S3, based on each traffic zone node The initial embedding vector and its neighboring traffic cell nodes The combined features of the initial embedding vectors are used to calculate the traffic cell nodes through a single-head attention mechanism. With each adjacent traffic zone node Attention scores between them;

[0010] S4, at the traffic zone node Within the set of neighboring nodes, the attention score is normalized to obtain the traffic cell node. For each adjacent traffic zone node Attention weights; based on the attention weights, neighboring traffic cell nodes... The initial embedding vectors are weighted and aggregated to form an intermediate feature vector containing neighborhood difference information;

[0011] S5. Construct an adjacency matrix with self-loops and normalize it based on the node degree matrix to obtain the structure propagation matrix; use the structure propagation matrix to perform graph convolution update on the intermediate feature vectors to obtain the traffic cell nodes. The final embedding vector forms the basic hybrid graph convolutional unit that combines single-head attention weighted aggregation with structure-normalized graph convolutional propagation;

[0012] Stack at least two layers of basic hybrid graph convolutional units to form a multi-layer hybrid graph convolutional neural network;

[0013] S6, Traffic Community Node The final embedding vector is used as the input feature for downstream tasks. The network parameters are trained using a loss function and backpropagation algorithm to achieve traffic zone modeling, travel destination prediction, or regional spatial feature analysis.

[0014] Preferably, in S1, the original feature vector includes at least one or more of the following types of features:

[0015] Spatial and morphological characteristics: Geometric center coordinates, area, perimeter of boundaries, and distance statistics to adjacent traffic areas of the traffic zone;

[0016] Land use and functional characteristics: the proportion or intensity of land use for residential, employment, commercial, and public services, or the distribution vector of POI categories;

[0017] Traffic supply characteristics: road density, road network hierarchy, density of bus or subway stations, density of transfer facilities, and parking supply indicators;

[0018] Travel attraction and activity intensity characteristics: historical arrivals, departures, OD interaction intensity, commuting time statistics or accessibility indicators;

[0019] Environmental characteristics: slow-moving accessibility, mixed-use index, population density, or nighttime light intensity.

[0020] As a preferred embodiment, in S2, the linear mapping employs a fully connected layer to unify the dimensionality of the traffic cell node attribute features. Specifically,

[0021] ;

[0022] For any traffic cell node Its original feature vector is denoted as The initial embedding vector obtained after linear transformation is represented as follows: ;

[0023] This is a learnable / trainable weight matrix; This is a trainable bias vector.

[0024] As a preferred embodiment, in S3, the specific formula for calculating the attention score is as follows:

[0025] ;

[0026] in, This is a trainable attention weight vector;

[0027] Traffic community nodes With adjacent traffic zone nodes Embedded concatenated vectors;

[0028] It is a non-linear activation function;

[0029] Attention score represents the traffic cell node. For adjacent traffic cell nodes The intensity of travel attraction.

[0030] As a preferred option, in S4, the attention score is normalized using the Softmax function;

[0031] The specific formula for calculating attention weights is as follows:

[0032] ;

[0033] in, Traffic community nodes For adjacent traffic cell nodes Attention weights; Traffic community nodes The neighbor set, representing the traffic cell node. Adjacent traffic zone nodes A set;

[0034] The specific formula for calculating the intermediate feature vector is as follows:

[0035] ;

[0036] in, The intermediate feature vector reflects the traffic cell nodes under the influence of differentiated neighborhoods. The aggregated representation is used for structure-normalized graph convolutional propagation updates.

[0037] As a preferred option, in S5, the method for constructing the structure propagation matrix is ​​as follows:

[0038] ;

[0039] The adjacency matrix of the original graph indicates whether there are physical connections or logical reachability relationships between traffic zones. Indicates traffic zone nodes and adjacent traffic zone nodes There are edges connecting them. Indicates no connection;

[0040] : Identity matrix, indicating the inclusion of self-loops, meaning that each traffic zone also considers its own information;

[0041] : A node degree matrix, where the diagonal lines represent the number of connections for each node. , This means placing this degree vector on the diagonal to form a diagonal matrix;

[0042] The structure propagation matrix is ​​symmetrically normalized;

[0043] Traffic Community Node The final embedding vector is specifically,

[0044] ;

[0045] in, It is a non-linear activation function; Traffic community nodes The final embedding vector.

[0046] Preferably, in S6, a supervised learning loss function is used to train the network parameters, including at least cross-entropy loss, mean squared error loss, or Huber loss; and the backpropagation algorithm is used to update the model parameters, including at least linear mapping parameters, attention parameters, and graph convolution propagation parameters.

[0047] As a preferred approach, attention weights are visualized and analyzed. The specific method is as follows:

[0048] For each traffic zone node Select the one with the highest attention weight Adjacent traffic zone nodes Form an attention connection set; represent traffic zones as nodes, and characterize them by the thickness or gray level of the connection lines. The size is used to generate an interactive visualization of the traffic area.

[0049] A hybrid graph convolutional neural network traffic cell modeling system incorporating attention mechanisms includes,

[0050] The graph data construction and preprocessing module is used to construct a graph structure with traffic zones as nodes and spatial adjacency or travel accessibility as edges, and to standardize or normalize the node features.

[0051] The feature mapping module is used to perform linear mapping on the original feature vectors of nodes to obtain the initial embedding vectors;

[0052] The single-head attention score calculation module is used to calculate traffic cell nodes. With adjacent traffic zone nodes Attention score;

[0053] The attention weight normalization module is used to normalize the attention score within the neighborhood to obtain the attention weight;

[0054] The weighted aggregation module is used to aggregate adjacent traffic cell nodes based on attention weights. Weighted aggregation is performed to obtain intermediate feature vectors;

[0055] The structure normalization propagation module is used to construct the structure propagation matrix and perform graph convolution propagation updates to obtain the final embedding vector.

[0056] The training and application module is used to train the system parameters based on the loss function and backpropagation, and to embed the node outputs for traffic zone modeling, travel destination prediction or regional spatial feature analysis.

[0057] The visualization and explanation module is used to output the attention weight matrix and the visualization results of traffic cell interactions.

[0058] Preferably, the hybrid graph convolutional neural network is embedded as a module in the multi-task travel prediction network structure and set in the region coding layer or destination candidate coding layer used to generate spatial representations. Its output node embedding represents the spatial input features as destination prediction, route selection or transportation mode decision-making tasks, so as to enhance the spatial sensitivity of the tasks.

[0059] The present invention has the following beneficial effects:

[0060] 1. The Hybrid Graph Convolutional Neural Network (HA-GCN) traffic cell modeling method and system proposed in this invention improves the neighborhood information aggregation method based on the traditional graph convolutional neural network structure. By introducing a learnable single-head attention mechanism, it achieves adaptive modeling of heterogeneous spatial relationships between traffic cells. Compared with existing graph convolutional models that use fixed weights or averaging aggregation methods, this invention can effectively distinguish the relative importance of different neighboring traffic cells in the spatial propagation process, thereby more accurately characterizing the potential travel attraction relationship and spatial interaction intensity between traffic cells.

[0061] 2. This invention, while maintaining the stability of the normalized propagation of the graph convolutional neural network structure, introduces a lightweight attention weight calculation mechanism. This enhances the model's ability to express regional functional differences, uneven travel attraction intensity, and weakly connected regions without significantly increasing computational complexity and storage overhead. By combining attention-weighted aggregation with GCN normalized propagation, it avoids the training instability and generalization performance degradation that may occur with pure attention models in large-scale traffic area maps, thus improving the model's robustness and scalability in real-world traffic data scenarios.

[0062] 3. The attention weight matrix generated by this invention has clear physical and behavioral meanings, and can be used to visualize and analyze the interaction strength between traffic zones. It intuitively reflects the differences in functional division and attractiveness of different areas within the urban transportation system, thus enabling interpretable output of the model results. This attention weight-based interpretation method helps transportation planners understand the model's decision-making logic, providing an intuitive basis for optimizing regional traffic organization, formulating travel guidance strategies, and adjusting the layout of transportation facilities.

[0063] 4. The HA-GCN module of this invention possesses excellent modularity and scalability. It can be used as an independent regional modeling method for traffic zone function identification, regional attractiveness assessment, and spatial interaction analysis. It can also be embedded in a multi-task travel prediction network structure as a spatial representation learning module, providing high-quality spatial feature inputs for tasks such as travel destination prediction, route selection, and mode of transport decision-making. Compared to existing technologies, this invention improves the accuracy and practicality of traffic zone modeling results while balancing model stability, computational efficiency, and interpretability.

[0064] 5. This invention introduces a single-head attention mechanism into the neighborhood aggregation stage of a graph convolutional neural network, constructing a traffic zone modeling method that is "stable in structure propagation, flexible in difference aggregation, and interpretable in results." This provides a new technical path for urban travel behavior analysis and transportation planning and management, and has significant practical application value for improving the refined analysis capabilities and intelligent decision-making level of urban transportation systems. Attached Figure Description

[0065] Figure 1 This is a schematic diagram of the overall process of the Hybrid Graph Convolutional Neural Network (HA-GCN) traffic cell modeling method that incorporates the attention mechanism of the present invention.

[0066] Figure 2 This is a schematic diagram of the structure of the HA-GCN module of the present invention.

[0067] Figure 3 This is a schematic diagram illustrating the embedding method of the HA-GCN module of the present invention in a multi-task travel prediction network structure. Detailed Implementation

[0068] The present invention will now be described in further detail with reference to the accompanying drawings and specific preferred embodiments.

[0069] In the description of this invention, it should be understood that the terms "left side," "right side," "upper part," "lower part," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. "First," "second," etc., do not indicate the importance of the components, and therefore should not be construed as a limitation of this invention. The specific dimensions used in this embodiment are only for illustrating the technical solution and do not limit the scope of protection of this invention.

[0070] like Figures 1-3As shown, a Hybrid Graph Convolutional Neural Network (HA-GCN) traffic cell modeling method incorporating attention mechanism is applicable to various scenarios such as traffic cell spatial representation learning, travel destination prediction, regional attractiveness assessment, traffic flow spatial feature extraction, and urban travel behavior analysis.

[0071] In this invention, a "traffic zone" refers to a basic spatial unit defined in urban traffic analysis and travel behavior modeling based on factors such as spatial continuity, road network structure, land use characteristics, and travel activity characteristics. Each traffic zone comprehensively represents the travel supply conditions, land use functional attributes, and travel activity intensity within a certain spatial range, and is a commonly used analytical unit in traffic planning and traffic demand analysis.

[0072] In the Hybrid Graph Convolutional Neural Network (HA-GCN) model that incorporates an attention mechanism, each traffic zone is abstracted as a node in a graph structure, and the spatial adjacency, road accessibility, or travel interaction relationships between traffic zones are abstracted as edges between nodes. This approach maps the urban spatial structure into graph-structured data, enabling unified modeling of the spatial dependencies and travel interaction characteristics of traffic zones through graph neural networks.

[0073] Appendix Figure 1 This diagram illustrates the overall workflow of a Hybrid Graph Convolutional Neural Network (HA-GCN) traffic cell modeling method that incorporates an attention mechanism. It showcases the main computational steps from graph data definition and preprocessing, node feature mapping and attention preparation, attention score calculation, attention weight normalization, neighborhood weighted aggregation, to structure-normalized graph convolutional propagation, as detailed below:

[0074] Step 1: Graph Data Definition and Preprocessing

[0075] Acquire traffic zone data for the target city, construct a graph structure with traffic zones as nodes and spatial adjacency or travel accessibility relationships as edges, forming graph data containing node sets and adjacency relationships, and standardize or normalize the node attribute features. In the graph data, nodes correspond to traffic zone spatial units, and edges correspond to spatial adjacency or travel accessibility relationships between traffic zones.

[0076] In the process of defining and preprocessing graph data, a corresponding original feature vector is constructed for each traffic cell node. The original feature vector includes at least one or more of the following types of features:

[0077] (1) Spatial and morphological characteristics: Geometric center coordinates, area, perimeter of the boundary, and distance statistics of the traffic area to adjacent traffic areas;

[0078] (2) Land use and functional characteristics: the proportion or intensity of land use for residential, employment, commercial, and public services, or the distribution vector of POI categories;

[0079] (3) Traffic supply characteristics: road density, road network hierarchy, density of bus or subway stations, density of transfer facilities, and parking supply indicators;

[0080] (4) Characteristics of travel attraction and activity intensity: historical arrivals, departures, OD interaction intensity, commuting time statistics or accessibility indicators;

[0081] (5) Other environmental characteristics: characteristics that characterize the activity level of a region, such as slow-moving accessibility, mixed-use index, population density, or nighttime light intensity;

[0082] The original feature vector is then standardized, normalized, or binned before being used as a node attribute input to the subsequent node feature mapping step.

[0083] Step 2: Node Feature Mapping and Attention Preparation

[0084] For each traffic zone node The original feature vectors are subjected to feature construction and linear mapping processing. The original feature vectors include at least node attribute features representing the spatial attributes, land use function attributes, and travel attraction characteristics of the traffic zone. Through linear transformation, the traffic zone node attribute features are mapped to a unified-dimensional embedding space to obtain the traffic zone nodes. The initial embedding vector, i.e. the initial embedding representation, provides the input feature basis for subsequent weighted aggregation of neighbor nodes based on the single-head attention mechanism.

[0085] Specifically, for any traffic cell node Its original feature vector is denoted as:

[0086] ;

[0087] A linear transformation is performed through a fully connected layer to obtain the initial embedding representation of the node, i.e., the initial embedding vector:

[0088] ;

[0089] in,

[0090] Learnable / trainable weight matrix (used to unify feature dimensions);

[0091] Trainable bias vector;

[0092] Traffic Community Node The initial embedding vector after feature mapping.

[0093] Step 3: Neighbor Node Attention Modeling

[0094] For any traffic zone node in the diagram Based on its own initial embedding vector and neighboring nodes (i.e., adjacent traffic cell nodes) The combined features of the initial embedding vectors are used to compute the node's values ​​through a single-head attention mechanism. With each neighboring node Attention scores are used to characterize the relative influence of different neighboring nodes on the node.

[0095] Specifically, for any pair of adjacent nodes Attention score The calculation is as follows:

[0096] ;

[0097] in, This is a trainable attention weight vector;

[0098] Traffic community nodes With adjacent traffic zone nodes The embedded concatenated vector is used to capture the joint features of the two;

[0099] It is a non-linear activation function; the single-head attention mechanism is used to reduce the computational and storage overhead of multi-head attention and keep the model lightweight;

[0100] Attention score represents the traffic cell node. For adjacent traffic cell nodes The intensity of travel attraction, measured from Departure to destination The intensity of travel attraction or preference (the higher the score, the more important it is).

[0101] The single-head attention mechanism is used to achieve differentiated modeling of the importance of neighboring nodes while maintaining the simplicity of the model structure, thereby reducing the computational and storage overhead caused by the multi-head attention mechanism.

[0102] Step 4: Attention weight normalization

[0103] At the node The set of neighboring nodes Within this process, the attention scores are normalized to obtain node values. For each neighbor node Differentiated attention weights.

[0104] Specifically, normalization is performed using the Softmax function:

[0105] ;

[0106] in:

[0107] :node neighboring nodes Normalized attention weights;

[0108] Traffic Community Node The set of neighboring communities (including areas that are spatially adjacent or accessible by transportation).

[0109] Step 5: Weighted Neighbor Information Aggregation

[0110] Based on the attention weights, the initial embedding vectors of neighboring nodes are weighted and aggregated to form intermediate node representations (intermediate feature vectors) that contain neighborhood difference information.

[0111] For nodes Its weighted aggregation representation is calculated as follows:

[0112] ;

[0113] in, The intermediate feature vector represents the node's behavior under the influence of differentiated neighborhoods. The comprehensive feature representation is used for subsequent graph convolutional propagation updates.

[0114] Step 6: Structure Normalized Graph Convolution Propagation

[0115] Based on the intermediate node representation, a structure normalization propagation mechanism for graph convolutional neural networks is introduced. An adjacency matrix with self-loops is constructed based on the adjacency relationships of the graph structure, and normalized according to the node degree matrix to obtain the structure propagation matrix. This structure propagation matrix is ​​used to maintain the stability of graph convolutional propagation and achieve structural smoothing. The structure propagation matrix is ​​used to update the node features (intermediate feature vectors) using graph convolution, thereby fusing the differential influence information of neighboring nodes while maintaining the stability of the graph convolutional network propagation, to obtain the node's output embedding representation (final embedding vector), forming a basic hybrid graph convolutional unit that combines single-head attention weighted aggregation with structure normalized graph convolutional propagation. By stacking at least two layers of this basic hybrid graph convolutional unit ("single-head attention weighted aggregation + structure normalized propagation"), a multi-layer hybrid graph convolutional neural network is formed.

[0116] Specifically, in the original adjacency matrix Introducing the identity matrix on this basis ,form And construct the degree matrix ,in The diagonal element is The sum of the elements in each row yields the symmetric normalized propagation matrix:

[0117] ;

[0118] The adjacency matrix of the original graph indicates whether there are physical connections or logical reachability relationships between traffic zones. Indicates traffic zone nodes and adjacent traffic zone nodes There are edges connecting them. Indicates no connection;

[0119] : Identity matrix, indicating the inclusion of self-loops, meaning that each traffic zone also considers its own information;

[0120] : A node degree matrix, where the diagonal lines represent the number of connections for each node. , This means placing this degree vector on the diagonal to form a diagonal matrix;

[0121] This is a symmetric normalized structure propagation matrix that enhances the stability of graph structure propagation. It primarily serves two purposes: ① smoothing node features: enabling information to propagate more evenly throughout the graph; ② mitigating gradient vanishing / exploding: making model training more stable by scaling the adjacency matrix.

[0122] The node features are updated using the structure propagation matrix. The output embedding is represented as:

[0123] ;

[0124] in, The activation function is nonlinear (such as ReLU); and a multi-layer HA-GCN network structure is formed by stacking at least two layers of hybrid graph convolutional layers of "single-head attention weighted aggregation + structure normalization propagation" to enhance the representation ability. Traffic community nodes The final embedding vector is used as the input feature for destination traffic zone selection.

[0125] Step 7: Network Training and Application

[0126] The node output embeddings are used as input features for downstream tasks. Supervised learning is employed to train the network parameters to achieve traffic zone modeling, travel destination prediction, or regional spatial feature analysis. During training, a loss function matching the downstream task is set, including but not limited to cross-entropy loss, mean squared error loss, or Huber loss. The model parameters, including at least linear mapping parameters, attention parameters, and graph convolution propagation parameters, are updated using a backpropagation algorithm.

[0127] Step 8: Visualization of the attention weight matrix

[0128] Attention weight matrix Perform visualization analysis, including:

[0129] (1) Select the node with the highest attention weight for each node. Each neighboring node forms an attention connection set;

[0130] (2) Using traffic zones as nodes and the thickness of the connecting lines or the shade of gray to represent the attention weight, generate a traffic zone interaction visualization map.

[0131] The above methods reveal the potential travel connections, functional divisions, and differences in attraction intensity between transportation zones.

[0132] Appendix Figure 3 This diagram illustrates the embedding of the HA-GCN module in a multi-task travel prediction network structure, demonstrating its application as a spatial representation learning module providing node embedding features for downstream travel prediction tasks. The HA-GCN module possesses strong modularity, allowing for flexible embedding into multi-task travel prediction network frameworks. Its output node embedding representations support applications such as travel destination prediction, traffic zone attractiveness assessment, spatial visualization, and functional zoning, enhancing the spatial sensitivity of tasks like travel destination prediction, route selection, and mode of transportation decision-making. This improves the model's ability to characterize and interpret spatial heterogeneity. Visual analysis of the attention weight distribution reveals the functional division and attraction intensity differences among different traffic zones in the urban travel system, achieving a unified approach from data-driven modeling to explanation of travel behavior mechanisms.

[0133] Based on the above method, this invention further proposes a Hybrid Graph Convolutional Neural Network (HA-GCN) traffic zone modeling system that incorporates an attention mechanism, comprising the following modules:

[0134] (1) Graph data construction and preprocessing module, used to construct a graph structure with traffic zones as nodes and spatial adjacency or travel accessibility as edges, and to standardize or normalize the node features;

[0135] (2) Feature mapping module, used to perform linear mapping on the original feature vectors of nodes to obtain the initial embedding representation;

[0136] (3) Single-head attention score calculation module, used to calculate the attention score between a node and its neighboring nodes;

[0137] (4) Attention weight normalization module, which is used to normalize the attention score in the neighborhood to obtain the attention weight;

[0138] (5) Weighted aggregation module, used to perform weighted aggregation of neighbor node embeddings based on attention weights to obtain intermediate node representations;

[0139] (6) Structure normalization propagation module, used to construct the structure propagation matrix and perform graph convolution propagation update to obtain the node output embedding representation;

[0140] (7) Training and application module, used to train system parameters based on loss function and backpropagation, and embed node outputs for traffic zone modeling, travel destination prediction or regional spatial feature analysis;

[0141] (8) Visualization Explanation Module, used to output attention weight matrix and traffic community interaction visualization results.

[0142] Appendix Figure 2 This diagram illustrates the structure of the HA-GCN module, showcasing a hybrid graph convolutional layer that combines a single-head attention mechanism with graph convolutional normalization propagation. The feature transformation module, attention score calculation module, attention weight normalization module, neighbor information aggregation module, and structure normalization propagation module are all implemented through neural network computation units. These modules are interconnected and exchange information via a directed computational graph structure, enabling node feature mapping, attention weight calculation, neighborhood information aggregation, and graph convolutional propagation to form differentiable computational paths. This supports end-to-end joint training based on the backpropagation algorithm, achieving unified optimization and automatic updating of model parameters.

[0143] The HA-GCN module proposed in this invention achieves a unified expression for heterogeneous modeling between traffic zones by integrating structured convolution and attention mechanisms. It provides a novel method for urban travel behavior analysis that combines predictive performance and interpretability. While balancing model computational efficiency and engineering deployability, it enhances the model's spatial sensitivity and interpretability. It is applicable to traffic zone modeling, travel destination prediction, and regional spatial behavior analysis in complex urban transportation systems, and can better support application scenarios such as traffic planning, travel guidance, and urban spatial optimization.

[0144] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A traffic zone modeling method incorporating a hybrid graph convolutional neural network with an attention mechanism, characterized in that, Includes the following steps: S1. Obtain traffic zone data for the target city and construct a graph structure with traffic zones as nodes and spatial adjacency or travel reachability relationships as edges. For each traffic zone node... Construct the corresponding original feature vectors and perform preprocessing to form a vector containing traffic zone nodes. Graph data including adjacency or travel relationships, and traffic community node attribute characteristics; S2, For each traffic zone node The original feature vectors are used for feature construction and linear mapping. Through linear transformation, the attribute features of traffic cell nodes are mapped to an embedding space of uniform dimension, thus obtaining the traffic cell node... The initial embedding vector; S3, based on each traffic zone node The initial embedding vector and its neighboring traffic cell nodes The combined features of the initial embedding vectors are used to calculate the traffic cell nodes through a single-head attention mechanism. With each adjacent traffic zone node Attention scores between them; S4, at the traffic zone node Within the set of neighboring nodes, the attention score is normalized to obtain the traffic cell node. For each adjacent traffic zone node Attention weights; Based on attention weights, adjacent traffic cell nodes The initial embedding vectors are weighted and aggregated to form an intermediate feature vector containing neighborhood difference information; S5. Construct an adjacency matrix with self-loops and normalize it based on the node degree matrix to obtain the structure propagation matrix; use the structure propagation matrix to perform graph convolution update on the intermediate feature vectors to obtain the traffic cell nodes. The final embedding vector forms the basic hybrid graph convolutional unit that combines single-head attention weighted aggregation with structure-normalized graph convolutional propagation; Stack at least two layers of basic hybrid graph convolutional units to form a multi-layer hybrid graph convolutional neural network; S6, Traffic Community Node The final embedding vector is used as the input feature for downstream tasks. The network parameters are trained using a loss function and backpropagation algorithm to achieve traffic zone modeling, travel destination prediction, or regional spatial feature analysis.

2. The hybrid graph convolutional neural network traffic cell modeling method with fused attention mechanism according to claim 1, characterized in that, In S1, the original feature vector includes at least one or more of the following types of features: Spatial and morphological characteristics: Geometric center coordinates, area, perimeter of boundaries, and distance statistics to adjacent traffic areas of the traffic zone; Land use and functional characteristics: the proportion or intensity of land use for residential, employment, commercial, and public services, or the distribution vector of POI categories; Traffic supply characteristics: road density, road network hierarchy, density of bus or subway stations, density of transfer facilities, and parking supply indicators; Travel attraction and activity intensity characteristics: historical arrivals, departures, OD interaction intensity, commuting time statistics or accessibility indicators; Environmental characteristics: slow-moving accessibility, mixed-use index, population density, or nighttime light intensity.

3. The hybrid graph convolutional neural network traffic cell modeling method with fused attention mechanism according to claim 1, characterized in that, In S2, the linear mapping uses a fully connected layer to unify the dimensionality of the traffic cell node attribute features. Specifically... ； For any traffic cell node Its original feature vector is denoted as The initial embedding vector obtained after linear transformation is represented as follows: ; This is a learnable / trainable weight matrix; This is a trainable bias vector.

4. The hybrid graph convolutional neural network traffic cell modeling method with fused attention mechanism according to claim 1, characterized in that, In S3, the specific formula for calculating the attention score is as follows: ； in, This is a trainable attention weight vector; Traffic community nodes With adjacent traffic zone nodes Embedded concatenated vectors; It is a non-linear activation function; Attention score represents the traffic cell node. For adjacent traffic cell nodes The intensity of travel attraction.

5. The hybrid graph convolutional neural network traffic cell modeling method with fused attention mechanism according to claim 1, characterized in that, In S4, the attention score is normalized using the Softmax function; The specific formula for calculating attention weights is as follows: ； in, Traffic community nodes For adjacent traffic cell nodes Attention weights; Traffic community nodes The neighbor set, representing the traffic cell node. Adjacent traffic zone nodes A set; The specific formula for calculating the intermediate feature vector is as follows: ； in, The intermediate feature vector reflects the traffic cell nodes under the influence of differentiated neighborhoods. The aggregated representation is used for structure-normalized graph convolutional propagation updates.

6. The hybrid graph convolutional neural network traffic cell modeling method with fused attention mechanism according to claim 1, characterized in that, In S5, the specific method for constructing the structure propagation matrix is ​​as follows: ； The adjacency matrix of the original graph indicates whether there are physical connections or logical reachability relationships between traffic zones. Indicates traffic zone nodes and adjacent traffic zone nodes There are edges connecting them. Indicates no connection; : Identity matrix, indicating the inclusion of self-loops, meaning that each traffic zone also considers its own information; : A node degree matrix, where the diagonal lines represent the number of connections for each node. , This means placing this degree vector on the diagonal to form a diagonal matrix; The structure propagation matrix is ​​symmetrically normalized; Traffic Community Node The final embedding vector is specifically, ； in, It is a non-linear activation function; Traffic community nodes The final embedding vector.

7. The hybrid graph convolutional neural network traffic cell modeling method with fused attention mechanism according to claim 1, characterized in that, In S6, supervised learning loss functions are used to train the network parameters, including at least cross-entropy loss, mean squared error loss, or Huber loss; the backpropagation algorithm is used to update the model parameters, including at least linear mapping parameters, attention parameters, and graph convolution propagation parameters.

8. The method for modeling traffic zones using a hybrid graph convolutional neural network with an attention mechanism as described in claim 1, characterized in that, The attention weights are visualized and analyzed using the following method: For each traffic zone node Select the one with the highest attention weight Adjacent traffic zone nodes Form an attention connection set; represent traffic zones as nodes, and characterize them by the thickness or gray level of the connection lines. The size is used to generate an interactive visualization of the traffic area.

9. A traffic zone modeling system incorporating a hybrid graph convolutional neural network with an attention mechanism, based on the modeling method described in any one of claims 1-8, characterized in that, include, The graph data construction and preprocessing module is used to construct a graph structure with traffic zones as nodes and spatial adjacency or travel accessibility as edges, and to standardize or normalize the node features. The feature mapping module is used to perform linear mapping on the original feature vectors of nodes to obtain the initial embedding vectors; The single-head attention score calculation module is used to calculate traffic cell nodes. With adjacent traffic zone nodes Attention score; The attention weight normalization module is used to normalize the attention score within the neighborhood to obtain the attention weight; The weighted aggregation module is used to aggregate adjacent traffic cell nodes based on attention weights. Weighted aggregation is performed to obtain intermediate feature vectors; The structure normalization propagation module is used to construct the structure propagation matrix and perform graph convolution propagation updates to obtain the final embedding vector. The training and application module is used to train the system parameters based on the loss function and backpropagation, and to embed the node outputs for traffic zone modeling, travel destination prediction or regional spatial feature analysis. The visualization and explanation module is used to output the attention weight matrix and the visualization results of traffic cell interactions.

10. The hybrid graph convolutional neural network traffic cell modeling system with fused attention mechanism according to claim 9, characterized in that, The hybrid graph convolutional neural network is embedded as a module in the multi-task travel prediction network structure and is set in the region encoding layer or destination candidate encoding layer used to generate spatial representations. Its output node embedding representation serves as the spatial input feature for destination prediction, route selection, or mode of transportation decision-making tasks, thereby enhancing the spatial sensitivity of the tasks.

Images