Traffic big data intelligent analysis method based on topology compression and feature dimension reduction

Abstract

The invention relates to the field of traffic big data intelligent analysis, and provides a traffic big data intelligent analysis method based on topology compression and feature dimension reduction, comprising the following steps: S1, obtaining time sequence traffic state data of each road segment node in a target road network, performing data cleaning and standardization processing, and dividing the data into a training set, a verification set and a test set; s2, on the basis of the training set data, the correlation degree between road segment nodes is calculated, and a road network topological structure including a node layer and a hyperedge layer is constructed; wherein each node uniquely belongs to one hyperedge; s3, on the basis of a road network topological structure, coding and compressing the node features by using a collective message passing mechanism to obtain node compression features and hyperedge compression features after dimension reduction; and S4, fusing the node compression feature and the hyperedge compression feature, inputting a prediction model based on a graph attention mechanism, and predicting the future traffic congestion state of the target road section. According to the method, the problems of high calculation load, characteristic redundancy and insufficient prediction precision of large-scale traffic data are solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis technology for traffic big data, specifically to an intelligent analysis method for traffic big data based on topology compression and feature dimensionality reduction. Background Technology

[0002] With the acceleration of urbanization, urban traffic congestion has become increasingly severe. Guiyang, as one of the cities with the most prominent traffic congestion in China, urgently needs real-time perception and accurate prediction of traffic conditions across its main, secondary, and branch road networks. Current traffic big data analysis faces two major pain points: First, the sheer scale of traffic data (each batch containing 3000+ road segment nodes and multi-dimensional temporal features) leads to high computational complexity and poor real-time performance when directly modeled, making it difficult to meet the immediate response requirements of traffic control. Second, traditional feature processing methods often focus on isolated node features, neglecting higher-order topological relationships within the road network, and lack targeted compression strategies, resulting in feature redundancy and insufficient model generalization ability.

[0003] In existing technologies, while graph neural networks (such as GCN and GraphSAGE) can capture node relationships, they do not efficiently compress road network topology, resulting in a large number of parameters and low computational efficiency. Traditional autoencoders only focus on feature dimensionality reduction and do not combine road network physical connectivity and feature correlation to construct a topological structure, which easily leads to the loss of key traffic information after compression. At the same time, existing loss functions mostly use a single MSE loss, which fails to take into account the reconstruction accuracy of core features and key attributes, resulting in large errors in downstream prediction tasks and making it difficult to meet the prediction accuracy requirements (error ≤ 1 level) for traffic management.

[0004] Therefore, there is an urgent need for a smart analysis method for traffic big data that balances topology compression efficiency, feature reconstruction accuracy, and prediction performance. Summary of the Invention

[0005] The core objective of this invention is to address the problems of computational complexity, feature redundancy, and insufficient prediction accuracy in existing traffic big data analysis. It proposes a traffic state analysis method that integrates topology compression and feature dimensionality reduction, which is suitable for efficient processing of urban road network traffic data, topology modeling, and short-term / medium-to-long-term congestion index prediction. It can be directly applied to practical business scenarios such as intelligent traffic management and traffic flow scheduling optimization.

[0006] To achieve the above objectives, the following technical solution is adopted: a traffic big data intelligent analysis method based on topology compression and feature dimensionality reduction, comprising the following steps: S1: acquiring time-series traffic state data of each road segment node in the target road network, and performing data cleaning and standardization processing, dividing the standardized dataset into training set, validation set and test set according to time order; S2: calculating the correlation degree between road segment nodes based on the training set data, and aggregating multiple nodes into hyperedges according to the correlation degree to construct a road network topology structure containing a node layer and a hyperedge layer; wherein, each node uniquely belongs to a hyperedge; S3: based on the road network topology structure, encoding and compressing the node features using the Set-Messaging Program (SetMP) mechanism to obtain dimensionality-reduced node compressed features and hyperedge compressed features; S4: fusing the node compressed features and the corresponding hyperedge compressed features, and inputting them into a prediction model based on graph attention mechanism to predict the traffic congestion state of the target road segment at least one future time point.

[0007] Further, in step S2, a road network topology structure containing a node layer and a hyperedge layer is constructed based on the training set data. Specifically, this includes: setting an upper limit p for the number of nodes allowed in a single hyperedge, and determining the total number of hyperedges K based on the total number of nodes N; mapping each node feature to a node embedding vector using a pre-trained encoder based solely on the node features of the training set; calculating the similarity between nodes based on the node embedding vectors to form a similarity matrix representing the degree of association between nodes; using the similarity matrix as input, using the K-Means clustering algorithm to divide all nodes into K disjoint sets, each set being defined as a hyperedge; and generating a fixed topology object based on the partitioned node set and hyperedge set. The topology object remains unchanged after construction, and the data processing of the validation set and test set reuses this topology object.

[0008] Furthermore, during clustering, if a generated hyperedge contains more than p nodes, then clustering is performed again within that hyperedge to ensure that each generated hyperedge contains no more than p nodes.

[0009] Furthermore, in step S2, constructing the road network topology also includes a topology attribute enhancement step, specifically including: attaching a dynamic attribute vector to each node in the topology, wherein the dynamic attribute vector includes road level information, historical congestion patterns, time factors, and spatial flow; and calculating a weight for each hyperedge in the topology, wherein the weight is determined based on the similarity between the nodes constituting the hyperedge.

[0010] Further, step S3, which involves encoding and compressing node features using the Set-Messaging (SetMP) mechanism, includes the following steps: Topology feature extraction step: Based on the road network topology, the traffic state temporal features and additional attribute information of the nodes are fused to generate node embedding features and hyperedge embedding features containing high-order associations; Topology feature compression step: Based on the node embedding features and hyperedge embedding features, preliminary dimensionality reduction and deep compression are performed through Set-Messaging, which includes a two-layer structure of nodes and hyperedges, to obtain node compressed features and hyperedge compressed features with reduced dimensionality.

[0011] Furthermore, the topological feature extraction step specifically includes: node embedding optimization: concatenating the traffic state temporal core features of the node with the dynamic attribute vector, and updating the node embedding features through neural network mapping; hyperedge embedding generation: for each hyperedge, performing permutation-invariant aggregation on the node embedding features of all nodes within it, and generating hyperedge embedding features through neural network mapping of the aggregation results.

[0012] Furthermore, the topological feature compression step specifically includes: preliminary dimensionality reduction: for each node, its core features, node embedding features, and the hyperedge embedding features of its unique hyperedge are fused together, and a low-dimensional preliminary compressed node feature is generated through a first neural network mapping; deep compression: for each hyperedge, its hyperedge embedding features, the mean of the core features of all nodes inside it, and the mean of the preliminary compressed node feature are fused together, and a hyperedge compressed feature is generated through a second neural network mapping.

[0013] Furthermore, the compression model corresponding to the aggregated message passing mechanism is obtained through training: in the first stage, the node embedding output by the encoder is used to assist in constructing the road network topology; in the second stage, the compression model is trained based on the fixed road network topology; and the training objective of the compression model is to reconstruct the core features and key road attributes of the original traffic state through the decoder, and the loss function used is a weighted loss function that assigns different weights to the reconstruction error of the core features and the reconstruction error of the attribute features.

[0014] Furthermore, the prediction model based on the graph attention mechanism adopts the GATv2 model, and the structure of the prediction model includes: an input layer, whose input is a fusion feature formed by concatenating the node compression feature with the compression feature of the superedge to which it belongs; an attention layer, which includes two GATv2 layers, used to perform dynamic attention aggregation on the fusion feature based on the graph structure constructed by the physical connection relationship of the road network; and an output layer, configured to output the predicted value of the traffic congestion index of the target road segment at multiple consecutive time points in the future.

[0015] Furthermore, the training process of the prediction model includes a) adopting a step-by-step training strategy, firstly fixing the parameters of the compressed model that have been trained in the SetMP mechanism; then, updating only the parameters of the prediction model; b) the training objective of the GATv2 model is to minimize the mean square error between the predicted congestion index and the actual value at multiple consecutive future time points.

[0016] Compared with the prior art, the present invention achieves the following beneficial effects:

[0017] 1. Excellent compression efficiency and significantly reduced computational complexity: The SetMP framework is used to achieve hierarchical compression of topological features, supporting two core compression ratio configurations of 1 / 3 and 2 / 3. After compression of 3000+ nodes in a single batch, the total feature dimension is significantly reduced. Combined with the freeze compression module and the training strategy of only fine-tuning the prediction layer, the computational efficiency is improved by more than 40% compared with the traditional GCN, meeting the needs of real-time prediction.

[0018] 2. Ensure that core information is not lost during compression: This invention designs a weighted MSE loss function that minimizes the original features and reconstructed features. By reasonably allocating the weights of core features (congestion index sequence) and key attributes (road level weight, historical congestion intensity), it ensures that core information that is crucial to the prediction task is not lost during compression.

[0019] 3. Outstanding predictive performance, meeting actual business needs: By integrating the SetMP compression module and the GATv2 model, a full-process framework of "topology compression - feature extraction - association capture - accurate prediction" is constructed, which can accurately predict the congestion index of core road sections within 20 minutes and branch roads within 1 hour. The MAE of the test set is as low as 0.72 (≤0.8 target value), which is 41% lower than the MAE of the traditional GCN model and 34% lower than GraphSAGE. The error is controlled within level 1, which is fully adapted to the decision-making requirements of traffic management in Guiyang City.

[0020] 4. Stable topology and strong generalization ability: A fixed topology object is constructed based on the training set data. The hyperedges are strictly non-intersecting and integrate the physical connectivity of the road network with the correlation of features. The validation set and test set reuse the topology to avoid data leakage. The prediction error of the model fluctuates by ≤5% on different time periods and different road segments, demonstrating excellent generalization ability.

[0021] 5. High practicality and low deployment cost: The method of this invention is adapted to differentiated sampling intervals of 5 minutes (core road sections) and 15 minutes (sub-roads), supports direct input of JSON structured traffic data, and the output results can directly support traffic control operations such as traffic light timing adjustment and congestion relief. It has low deployment cost and strong feasibility.

[0022] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0023] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0024] Figure 1 This is a flowchart illustrating the intelligent analysis method for traffic big data based on topology compression and feature dimensionality reduction according to an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of the road network topology (nodes and hyperedges) in this embodiment of the invention and the traditional road network topology;

[0026] Figure 3 This is a schematic diagram of the feature extraction process of the SetMP (Set Messaging) mechanism in an embodiment of the present invention;

[0027] Figure 4 This is a schematic diagram of the topology feature compression process in an embodiment of the present invention;

[0028] Figure 5 This is the GATv2 prediction model structure in this embodiment of the invention;

[0029] Figure 6 This is a schematic diagram of the platform interface according to an embodiment of the present invention;

[0030] Figure 7 This is a performance comparison chart of the algorithm of this invention and three existing algorithms. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0033] Figure 1 This is a flowchart illustrating an intelligent analysis method for traffic big data based on topology compression and feature dimensionality reduction, according to an embodiment of the present invention. Figure 1 As shown, a method for intelligent analysis of traffic big data based on topology compression and feature dimensionality reduction includes the following steps:

[0034] S1: Obtain the time-series traffic status data of each road segment node in the target road network, and perform data cleaning and standardization processing. Divide the standardized dataset into training set, validation set and test set in time order.

[0035] Step S1 is used to acquire and preprocess traffic big data. The quality of traffic big data directly determines the accuracy of subsequent topology modeling and analysis. This step S1 uses a standardized data collection and cleaning process to clarify the feature dimensions, providing a high-quality, structured input data foundation for the entire technical solution. At the same time, it completes the time division of the dataset to avoid data leakage in subsequent topology construction and model training.

[0036] S1.1: Traffic Data Collection

[0037] Data source: Traffic data is collected periodically through the Guiyang Municipal Transportation Administration's Intelligent Transportation System and Baidu Maps Open Platform API, covering the main roads, secondary roads, and branch roads network throughout Guiyang City.

[0038] Data type and format: The collected data includes road segment ID, road segment center point latitude and longitude (WGS84 coordinate system), timestamp, average driving speed, congestion index (0-10 level), and road level label; the data format is JSON structured data, and the time sampling interval is set to 5 minutes (core road segment) and 15 minutes (sub-road); at the spatial level, the road segment location can be associated with a 50m × 50m resolution raster grid (only used for spatial range matching, the core analysis still uses the road segment as the basic unit).

[0039] Data Scale and Feature Dimensions: A single batch of collected data includes Each road segment node, time window length (That is, each node has a feature dimension of 10, corresponding to a congestion index sequence of 10 consecutive timestamps; average driving speed is used as a supplementary node attribute and is not included in the core feature dimension), forming a set of sub-signals. ,in .

[0040] The total number of road segment nodes, included in a single batch of collected data. A road section node in Guiyang City (core road section + branch road). The core feature dimension of a single node. This corresponds to a congestion index sequence of 10 consecutive timestamps (average driving speed is a node attribute supplement and is not included in this dimension). : The set of traffic sub-signals, including Each independent sub-signal unit covers traffic data collected in multiple batches / time periods. :gather The first in Each sub-signal corresponds to the core feature data of all nodes within a certain time period. : No. The first of the sub-signals The core feature vector of each node has a dimension of . The range of values ​​is the real number field. This represents the congestion index sequence of 10 consecutive timestamps for that node. A 3D real space is used to define the core feature vector of a node. The value space of .

[0041] S1.2: Data Cleaning and Standardization

[0042] (1) Handling outliers and missing values: using Principle of exclusion Outliers within the range; missing values ​​are interpolated using K-nearest neighbors (KNN). Data is filled in based on concurrent data of spatially adjacent road segments, and invalid data in non-road areas is directly filtered out. The principle is the outlier screening rule, among which... This is the average of historical traffic data (such as congestion index and speed) for a certain road segment. This represents the standard deviation of historical data for this road section; exceeding... Data within the specified range is identified as outliers and removed. : The number of neighboring nodes in the K-nearest neighbor interpolation This refers to referencing the contemporaneous data of the three spatially adjacent road segments when filling in missing values ​​for that node.

[0043] (2) Standardization processing: Min-max normalization is performed on traffic parameters (speed, congestion index) of different magnitudes, and the normalized data range is mapped to Min-max normalization is a data standardization method that maps traffic parameters of different orders of magnitude, such as speed and congestion index, to... Intervals eliminate dimensional differences.

[0044] (3) Dataset partitioning: To avoid data leakage during subsequent topology construction and model training, after data preprocessing, the standardized dataset is partitioned into a training set (first set, second set, and third set) according to time sequence. Time series data), validation set (intermediate) ), test set (after) The number of sub-signals contained in the training set after partitioning. Validation set test set Furthermore, there is no time overlap among the datasets.

[0045] in, The number of sub-signals contained in the training set, corresponding to the first... The total number of sub-signals in time series data. The number of sub-signals contained in the validation set, corresponding to the intermediate... The total number of sub-signals in time series data. The number of sub-signals contained in the test set, corresponding to the following... The total number of sub-signals in time series data.

[0046] The output forms a standardized traffic state dataset with partitioned elements. The training set, validation set, and test set all contain node feature matrices. Road segment attribute table (road grade, length, speed limit, average driving speed) and timestamp index.

[0047] Node feature matrix, dimension 1 , row corresponding Each road segment node, corresponding to the column The matrix is ​​a core feature and is included in the training set, validation set, and test set. OK The real matrix space of columns, used to define the node feature matrix. The dimensions and value space of .

[0048] S2: Based on the training set data, calculate the correlation degree between road segment nodes, and aggregate multiple nodes into hyperedges according to the correlation degree to construct a road network topology structure containing a node layer and a hyperedge layer; wherein, each node uniquely belongs to a hyperedge;

[0049] Step S2 is used for structural modeling based on road network topology. The standardized data that has been preprocessed and partitioned already has structured characteristics. This step S2 only constructs the road network topology structure and enhances its attribute information based on the training set data. Disjoint hyperedges are generated through a rigorous clustering algorithm to ensure that each node belongs to only one hyperedge, providing a structured foundation for subsequent topology compression.

[0050] Furthermore, in step S2, a road network topology structure containing a node layer and a hyperedge layer is constructed based on the training set data. This includes: setting an upper limit p for the number of nodes allowed in a single hyperedge, and determining the total number of hyperedges K based on the total number of nodes N; mapping each node feature to a node embedding vector using a pre-trained encoder based solely on the node features of the training set; calculating the similarity between nodes based on the node embedding vectors to form a similarity matrix representing the degree of association between nodes; using the similarity matrix as input, using the K-Means clustering algorithm to divide all nodes into K disjoint sets, each set being defined as a hyperedge; and generating a fixed topology object based on the partitioned node set and hyperedge set. The topology object remains unchanged after construction, and the data processing of the validation set and test set reuses this topology object. More specifically, this is achieved through the following steps:

[0051] S2.1: Road Network Topology Construction

[0052] This invention aims to construct a simplified topology that can simultaneously reflect the physical connectivity and traffic state similarity of a road network. To this end, it abandons the traditional graph structure that only includes simple edge connections and introduces "hyperedges" as higher-order structural units, such as... Figure 2 As shown, hyperedges capture hidden functional areas or congestion propagation clusters in the road network by aggregating nodes with similar features, thus providing richer structural information for subsequent compression and prediction.

[0053] S2.1.1: Setting the number and length of super-edges

[0054] Define the topology compression ratio: The topology compression ratio is the ratio of the total feature dimension of all nodes and hyperedges after compression to the total feature dimension of all nodes before compression. The topology compression ratio is adjusted by configuring the maximum hyperedge length, the compressed feature dimension of nodes, and the compressed feature dimension of hyperedges. Details are as follows:

[0055] Set the maximum node capacity of a single hyperedge (i.e., the maximum length of the hyperedge). Maximum length of the superedge. This is a key preset parameter that is directly related to the desired final feature compression ratio. By adjusting the value of p, the aggregation granularity of the system over the original road network topology can be controlled: a smaller... Value (e.g.) =6) will generate more and smaller hyperedges, which is beneficial for preserving more detailed local features and corresponds to a higher compressed feature dimension (such as compression ratio). Larger Value (e.g.) This results in coarser-grained aggregation, corresponding to a lower compressed feature dimension (e.g., topological compression ratio). Those skilled in the art can adaptively adjust the calculation efficiency and prediction accuracy according to the actual business requirements. The possible values ​​for . Total number of hyperedges (number of clusters). Based on the total number of nodes With preset The value is determined, for example, through a formula. Calculate to ensure all Each node can be assigned to a unique superedge; this is also the target number of clusters for K-Means clustering. This is an up-rounding function used to calculate the total number of superedges and the number of local K-Means sub-clusters, ensuring that no nodes are omitted from the allocation.

[0056] Maximum length of super edge The selection of compression efficiency and information retention needs to be balanced. A value that is too small will result in an excessive number of superedges, limiting the benefits of compression; An excessively high value will cause significant differences between nodes inside the hyperedge, resulting in the loss of local features. Typically, The value can be selected within a range based on the total number of nodes N and the desired compression ratio, for example... The value can range from 4 to 12. As an implementable configuration, when the desired compression ratio is... When it is 1 / 3, it can be taken as 1 / 3. =8; when expected When it is 2 / 3, it can be taken as =6.

[0057] S2.1.2: Node Embedding Pre-training

[0058] Initialize a 2-layer MLP encoder based solely on the node features of the training set. (Input layer) Hidden layer 64, output layer 20 (with ReLU activation function) to "reconstruct the core features of the nodes". (10-dimensional congestion index sequence) is used as the unsupervised target for pre-training—the core features of the j-th node are used as the target. Encoded as an embedding The original features are then reconstructed using a single MLP decoder layer with MSE as the loss function. The pre-training process lasts 50 epochs, using the Adam optimizer with a learning rate of [missing information]. This is used to optimize the parameters of the MLP encoder and complete the node embedding pre-training.

[0059] Layer MLP encoder (parameters are) Initialization is based solely on training set node features, with the input layer dimension... The system consists of 64 hidden layers and 20 output layers, with ReLU activation function used to encode node features into high-dimensional embeddings. : No. The core feature vector (10-dimensional congestion index sequence) of each node is the input of the MLP encoder. : No. The initial embedding vectors output by the MLP encoder for each node have a dimension of 20 and satisfy the following conditions: . In a 3D real space, define the initial embedding vector of the node. The value space of . MSE: Mean Squared Error Loss Function, used as an unsupervised objective in node embedding pre-training to measure the difference between the reconstructed node core features and the original features.

[0060] S2.1.3: Similarity Matrix Calculation

[0061] Pairwise similarity matrix: Also known as the "pairwise node similarity matrix," it's a square matrix representing the pairwise similarity between all nodes; specifically, it's the pairwise similarity matrix of node embeddings. Each element in the matrix corresponds to the similarity of only one pair of nodes, comprehensively covering the similarity relationships of all node combinations without omission or repetition. This embodiment of the invention calculates the pairwise similarity matrix based on the node embeddings output by the pre-trained encoder. The SNR distance metric is used to preserve semantic and feature relevance, as shown in the following formula:

[0062]

[0063] : The pairwise similarity matrix of node embeddings, with dimensions of Row / column correspondence Each node, element Representation Nodes With nodes Embedding similarity. OK In the space of real matrices of columns, define the similarity matrix. The dimensions and value space of . Similarity matrix The Middle Line number The elements of a column, i.e., nodes With nodes The embedding similarity is determined by the SNR distance metric function. Calculated. SNR (Short-Rate Measure) is a distance metric function used to calculate the similarity between the embeddings of two nodes. Its core logic is... . : are nodes respectively ,node The initial embedding vector (20-dimensional) is the input to the SNR distance metric function. :node With nodes The Euclidean distance (L2 norm) of the initial embedding vectors represents the degree of difference between the embedding vectors. : Noise coefficient of SNR distance metric, with a value of 0.1, used to balance the sensitivity of similarity calculation.

[0064] S2.1.4: K-Means clustering generates disjoint superedges

[0065] With similarity matrix As input, perform K-Means clustering (number of clusters) Each cluster is defined as a hyperedge. If the number of nodes in a certain cluster exceeds Then, perform local K-Means (sub-cluster count) again within the cluster. Number of cluster nodes / ), ensure all extra-edge lengths ; final generation A strictly disjoint set of hyperedges Each node belongs to one and only one hyperedge. K-Means is a clustering algorithm that uses a similarity matrix... Given the input, the number of clusters is... This is used to generate disjoint hyperedge clusters; if the number of nodes in the cluster exceeds... If so, then local K-Means subdivision is performed.

[0066] in, A single hyperedge (cluster) is the basic unit of a higher-order topology, containing several road segment nodes, with the number of nodes not exceeding [a certain threshold]. .

[0067] The final topology object output is a fixed topology object containing nodes and hyperedges. (No need to build basic edges) (The hyperedge has integrated road segment connectivity and correlation information). This topology is built only based on the training set. Subsequent validation and test set data reuse this topology structure and do not participate in topology updates.

[0068] Fixed topology object, For a set of nodes ( (each road segment node) This is a set of superedges, constructed solely based on the training set; the validation / test set reuses this topology. : Hyperedge set, containing Each node belongs to only one of the strictly disjoint superedges. : A set of road segment nodes, containing Each road segment node in Guiyang City is a basic unit of the topology object.

[0069] S2.2: Enhanced Topology Attributes

[0070] To enhance the ability of the topology to represent traffic conditions, targeted attribute information is added to nodes and hyperedges to strengthen the physical meaning of relationships. Topology attribute enhancement specifically includes: adding a dynamic attribute vector to each node in the topology, which contains road level information, historical congestion patterns, time factors, and spatial flow; and calculating a weight for each hyperedge in the topology, with the weight determined based on the similarity between the nodes constituting that hyperedge. This is achieved through the following steps:

[0071] S2.2.1: Attaching Node Attributes

[0072] Each node is appended with a dynamic attribute vector. (with timestamps) renew), :node timestamp The dynamic attribute vector below, with dimension 1 With timestamp It is updated in real time and includes four types of attributes: road level weight, historical congestion intensity, time weight, and spatial weight. A 3D real space is defined to represent the dynamic attribute vectors of nodes. The value space of . Timestamp: Represents the time node for traffic data collection. The sampling interval is 5 minutes for core road sections and 15 minutes for branch roads. Node dynamic attribute vector. It updates synchronously with this timestamp. As an example, each node is appended with a dynamic attribute vector. Include:

[0073] ① Road grade weighting (3 for main roads, 2 for secondary roads, and 1 for local roads);

[0074] ② Historical congestion intensity (average congestion index for the same period in the past 30 days within the training set);

[0075] ③ Time weight (dynamically adjusted according to timestamp: peak hours 1.2, off-peak hours 1.0, low-peak hours 0.8);

[0076] ④ Spatial weights (normalized values ​​of traffic flow on road segments during the same period in the training set).

[0077] All attributes are incorporated into the node embedding update through feature concatenation, where ① and ② are the core attributes that are crucial for prediction, and ③ and ④ are auxiliary and enhancing attributes.

[0078] S2.2.2: Add superedge attribute

[0079] The weight of a hyperedge is determined by the average similarity of all nodes within that hyperedge, as shown in the formula: This enhances the correlation characterization of higher-order structures.

[0080] : Single hyperedge The weight value is calculated from the average similarity among all nodes within the hyperedge, and is used to enhance the relevance representation of the higher-order structure of the hyperedge. Mean function, used here to calculate hyperedge The average similarity of all pairs of nodes within the hyperedge, with the input being the hyperedge. Any two nodes similarity . :node With nodes Embedding similarity (taken from pairwise similarity matrix) (), is the basic unit for calculating hyperedge weights. : Constraints, representing nodes and nodes All belong to the current superedge Only the mean similarity between nodes inside the superedge is calculated. : A single hyperedge (cluster), which is used here as the object for calculating the hyperedge weight.

[0081] S3: Based on the road network topology, the features of the nodes are encoded and compressed using the Set-Messaging Program (SetMP) mechanism to obtain the dimensionality-reduced compressed features of the nodes and the compressed features of the hyperedges.

[0082] Step S3 is used to implement topological feature extraction and topological feature compression based on Set-based Message Passing (SetMP). Based on the fixed topological objects constructed from the training set, step S3 first transforms isolated node features into topological features containing higher-order relationships, and then implements hierarchical topological compression (including feature dimensionality reduction) through the SetMP architecture. This clarifies the phased training logic of the compression model, reducing computational complexity while preserving the core spatiotemporal correlations.

[0083] Further, step S3, using a collective message passing mechanism, encodes and compresses the node features, including the following steps: Topological feature extraction step: Based on the road network topology, the traffic state temporal features and additional attribute information of the nodes are fused to generate node embedding features and hyperedge embedding features containing high-order associations; Hierarchical topological feature compression step: Based on the node embedding features and hyperedge embedding features, preliminary dimensionality reduction and deep compression are performed through a collective message passing mechanism containing two layers of nodes and hyperedges to obtain node compressed features and hyperedge compressed features with reduced dimensionality. More specifically, this includes the following steps:

[0084] S3.1: Topological Feature Extraction

[0085] Step S3.1 transforms isolated node numerical features into topological features containing higher-order correlations, laying the foundation for subsequent compression. Further, step S3.1 includes: node embedding optimization: concatenating the core traffic state time-series features of nodes with dynamic attribute vectors, and updating the node embedding features through neural network mapping; hyperedge embedding generation: for each hyperedge, performing permutation-invariant aggregation on the node embedding features of all nodes within it, and generating hyperedge embedding features through neural network mapping of the aggregation results. For example... Figure 3 As shown, this is achieved through the following steps:

[0086] S3.1.1: Node Embedding Optimization

[0087] Core features of nodes With dynamic attribute vectors The nodes are concatenated and updated using a 2-layer MLP to retrieve the pre-trained node embeddings. After the update The MLP activation function is ReLU to avoid gradient vanishing.

[0088] : No. The core feature vector of each node is the basic input for node embedding optimization. :node timestamp The dynamic attribute vector below, with dimension 1 It includes four types of attributes, such as road level weight and historical congestion intensity, which are combined with the core features of the node to participate in the embedding optimization. : No. The initial embedding vector (pre-trained output) or updated node embedding vector for each node, with dimension preserved after optimization. It contains the core features and attribute information of the node. In a 3D real space, define the node embedding vector. The value space of . ReLU: Rectified Linear Unit activation function, used in a 2-layer MLP for node embedding optimization to avoid the gradient vanishing problem and enhance the nonlinear expressive power of the model.

[0089] S3.1.2: Hyperedge Embedding Generation

[0090] Hyperedge embeddings are generated using a permutation-invariant function (concatenation of mean, maximum, and minimum values), ensuring that the result is unaffected by the node order. The formula is as follows:

[0091]

[0092] ReLU: Rectified Linear Unit Activation Function, used here to... After the output, the nonlinear expressive power of the model is enhanced, and gradient vanishing is avoided. : Super-edge The initial embedding vector, with dimension . It is generated by concatenating the mean, maximum and minimum values ​​embedded in the nodes inside the hyperedge and then processing them through MLP, containing high-order correlation information of the hyperedge. Feature concatenation operation, used to combine the core features of nodes. With dynamic attribute vectors Concatenation, or concatenation of mean, maximum, and minimum features during superedge embedding generation. Layer MLP (parameters are) ), used for hyperedge embedding generation, with an input dimension of 60 (20×3), corresponding to the concatenation dimension of the mean, maximum, and minimum values ​​of the node embeddings), an output dimension of 20, and the activation function being ReLU. Super-edge Embedded vectors of all nodes The mean of the permutation invariant function is used to ensure that the hyperedge embedding is not affected by the node arrangement order. : Super-edge Embedded vectors of all nodes The maximum value, the components of the permutation-invariant function, captures the extreme value features of the embedding of nodes inside the hyperedge. : Super-edge Embedded vectors of all nodes The minimum value of the permutation invariant function is used to capture the extreme value features of the embedding of nodes inside the hyperedge. : Constraints, representing nodes Belongs to the current superedge Statistical calculations are performed only on the embedding vectors of nodes inside the hyperedge. A single hyperedge (cluster) is an object generated by hyperedge embedding, corresponding to a fixed topology object. The super-edge unit in the middle.

[0093] S3.1.3: Feature Mapping Output

[0094] The original 10-dimensional node core features are mapped to 20-dimensional topological features containing higher-order relationships, realizing the transformation from isolated node features to topologically related features.

[0095] S3.2: SetMP-based topology feature compression

[0096] SetMP (Set-based Message Passing) is a pure set-based topological message passing framework that operates only on nodes and hyperedges (each node belongs to only one hyperedge). It achieves efficient compression of topological features through a hierarchical strategy of "preliminary dimensionality reduction + deep compression", and the compression model and hyperedge inference module are trained in stages.

[0097] Furthermore, the topological feature compression steps specifically include: preliminary dimensionality reduction: for each node, its core features, node embedding features, and the embedding features of its unique hyperedge are fused, and a low-dimensional preliminary compressed node feature is generated through a first neural network mapping; deep compression: for each hyperedge, its hyperedge embedding features, the average of the core features of all internal nodes, and the average of the preliminary compressed node features are fused, and a hyperedge compressed feature is generated through a second neural network mapping. For example... Figure 4 As shown, the specific implementation process is as follows:

[0098] S3.2.1: Preliminary Dimensionality Reduction (Superedge to Node Compression)

[0099] Feature dimensionality reduction is achieved by leveraging the association between hyperedge embeddings and node embeddings, as shown in the formula:

[0100]

[0101] :node The initial compressed feature vector has a dimension of It is generated by the compression process from hyperedge to node, realizing the dimensionality reduction of 20-dimensional topological features to low-dimensional features. Layer MLP (parameters are) This is used for initial dimensionality reduction from hyperedge to node, with an input dimension of 50 and an output dimension of 50. (Corresponding compression ratio) or (Corresponding compression ratio) The activation function is ReLU. :node The core feature vector has a dimension of 10 and corresponds to a congestion index sequence of 10 consecutive timestamps. Feature concatenation operation, used to combine the core features of nodes. Node embedding Hyperedge embedding Segmented by dimension. :node The updated embedding vector has dimensions of It contains the core features and attribute information of the node. :node Belongs to the only super edge The initial embedding vector, with dimension . This is one of the inputs for the initial dimensionality reduction. The dimension of the initial compressed feature of the node, taking a value of 2 or 5, determined by the topology compression ratio. Decide. Topological compression ratio: The ratio of the compressed feature dimension to the original feature dimension. correspond correspond In a real space, define the initial compression features of nodes. The value space of .

[0102] S3.2.2: Deep Compression (Node-to-Hyperedge Compression)

[0103] Based on the node features obtained from the initial dimensionality reduction, the hyperedge layer is deeply compressed to supplement the global association information of the hyperedges. The formula is as follows:

[0104]

[0105] : Super-edge The deep compressed feature vector, with dimension It is generated by the depth compression process from node to hyperedge, supplementing the global association information of the hyperedge. layer The first layer is used for feature mapping from node to hyperedge depth compression, with an input dimension of 32, a hidden layer dimension of 64, and the activation function is ReLU. layer The second layer is used for feature mapping from nodes to hyperedge depth compression, with an input dimension of 64 and an output dimension of 10. ReLU: Rectified Linear Unit activation function, applied here to the first layer of the MLP. After the output, the nonlinear expressive power of the model is enhanced. Mean aggregation function, for hyperedge The average value of the corresponding features of all nodes within the edge is taken to solve the problem of differences in the number of nodes on different hyperedges. : Super-edge Core features of all nodes The mean value, with a dimension of 10, is one of the inputs for deep compression. : Super-edge Preliminary compression features of all nodes within The mean, with a dimension of 2 (corresponding to) ), is one of the inputs for deep compression. : The dimension of the hyperedge depth compression feature, with a fixed value of 10. In a 3D real space, define the hyperedge depth compression feature. The value space of . A single road segment node is one of the basic operational units of the SetMP framework and belongs to only one superedge. A single hyperedge (cluster) is one of the basic operational units of the SetMP framework, containing several road segment nodes. : Constraints, representing nodes Belongs to the current superedge .

[0106] S3.2.3: Compression Ratio Calculation The overall compression ratio of topology compression is determined by the ratio of the total dimension of the node compression features and the hyperedge compression features to the total dimension of the original features, as shown in the formula:

[0107]

[0108] Example: , , , , .

[0109] Topological compression ratio: Represents the proportion of the total dimension of node and hyperedge features after compression to the total dimension of the original node features. The core configuration is... or . The total number of road segment nodes, included in a single batch of collected data. A road section node in Guiyang City. The node initially compresses the dimension of the feature, taking a value of 2 (corresponding to...). ) or 5 (corresponding) ). The total number of hyperedges is calculated using the following formula: (Round up), in the example hour . : The dimension of the hyperedge depth compression feature, with a fixed value of 10. The original core feature dimension of a single node is fixed at 10 (corresponding to a congestion index sequence of 10 consecutive timestamps). : The floor function, used to round up the total number of hyperedges. The calculation ensures that all nodes can be assigned to a unique superedge. Maximum length of the superedge, a key parameter. correspond correspond .

[0110] S3.2.4: Compressed Model Training (Phased)

[0111] The compressed model corresponding to the aggregated message passing mechanism is obtained through training. The training of the compressed model consists of two stages: "pre-trained encoder (hyperedge inference) + training compressed model," ensuring logical consistency between topology construction and compressed training. In the first stage, the encoder is pre-trained, and the node embeddings output by the encoder are used to assist in constructing the road network topology. In the second stage, the compressed model is trained based on the fixed road network topology. Furthermore, the training objective of the compressed model is to reconstruct the core features of the original traffic state and key road attributes through the decoder. The specific implementation is as follows:

[0112] S3.2.4.1: Phase 1 (Completed)

[0113] Pre-trained MLP encoder Used to generate node embeddings and build a fixed topology .

[0114] Layer MLP encoder (parameters are) The pre-training phase is used to generate node embeddings, providing a foundation for fixed topology construction. Fixed topology object, For a set of nodes, It is a set of superedges, constructed only based on the training set, and the validation / test set is reused.

[0115] S3.2.4.2: Phase 2 (Compressed Model Training)

[0116] 1. Training data: Only sub-signals from the training set are used. and fixed topology objects The validation set is used to monitor the training effect, while the test set is not used in training.

[0117] : Core feature data of all nodes within a certain time period, in the form of . : No. The first of the sub-signals The original core feature vector of each node, with a dimension of 10.

[0118] 2. Training objective: With the goal of "reconstructing the original core features + key attributes" (while also taking into account downstream prediction needs), reconstruct the 10-dimensional congestion index sequence and 2-dimensional core attributes (road level weight, historical congestion intensity) of the nodes.

[0119] 3. Training configuration:

[0120] Optimizer: Adam Optimizer (learning rate) Numerical stability parameters (Weight decay is set to 0, no regularization is required); where Adam: adaptive moment estimator optimizer, used to compress model training. : Learning rate, a parameter of the Adam optimizer, controls the step size for updating model parameters, and is set to 0.003 in this embodiment of the invention. : Numerical stability parameter, a parameter of the Adam optimizer, used to avoid a denominator of 0, and in this embodiment of the invention, the value is 0.001.

[0121] Training rounds: 300 rounds, batch size: 25.

[0122] Early stopping strategy: A termination strategy for model training. If the validation set reconstruction error (MSE) does not decrease for 10 consecutive rounds, training is stopped and the optimal model is saved. MSE: Mean Squared Error Loss Function, used to measure the difference between the reconstructed features and the original features, and is also the monitoring metric for the early stopping strategy.

[0123] 4. Loss Function: The weighted mean squared error (MSE) of the original and reconstructed features is minimized. This loss function uses MSE to differentiate and retain information crucial for downstream prediction tasks during feature compression. Different weights are assigned to the reconstruction errors of core traffic state features (e.g., congestion index sequences) and key road attribute features (e.g., road grade weights, historical congestion intensity). Let the weight of the core features be α, and the weight of the attribute features be β (satisfying α + β = 1). This is achieved through a decompression layer. (2-layer MLP, input dimension) Output dimension 12) Reconstructed features - In the output 12-dimensional vector, the first 10 dimensions correspond to the reconstructed core features. The last two dimensions correspond to the core attributes of the reconstruction. In a preferred embodiment, α = 0.8 and β = 0.2 are chosen to emphasize that the model prioritizes the fidelity of dynamic traffic state information during compression. This is achieved through a decompression layer. Reconstructing features, its loss function for:

[0124]

[0125] The weighted mean squared error loss function of the compressed model is derived from the core feature reconstruction loss (weights). ) and core attribute reconstruction loss (weight) It consists of two parts. : 2-layer MLP decompression layer, input dimension is (correspond The output dimension is 12, which is used to reconstruct the original core features and key attributes from the compressed features. :node In the The original 10-dimensional congestion index sequence (core feature) in each sub-signal. Nodes reconstructed from the decompression layer No. A 10-dimensional congestion index sequence of individual signals. :node In the Sub-signal timestamp The two core attributes include road level weight and historical congestion intensity. Nodes reconstructed from the decompression layer No. Sub-signal timestamp The 2D core attributes below. : The square of the L2 norm of the vector, used to calculate the difference between the original feature / attribute and the reconstructed feature / attribute. The number of sub-signals contained in the training set, corresponding to the first... The total number of sub-signals in time series data. : A set of road segment nodes, containing Each road segment node is the basic unit of a topology object. : Constraints, representing nodes Belongs to the set of nodes .

[0126] Typically, α should be greater than β; for example, α can take values ​​in the range of 0.6 to 0.9, and β correspondingly in the range of 0.4 to 0.1. In a specific and effective embodiment, α = 0.8 is used. =0.2, to emphasize that the model prioritizes the fidelity of dynamic traffic state information during compression. Those skilled in the art can adjust the above weight ratio according to the characteristics of different urban road networks or the focus of the prediction task.

[0127] To ensure that the compressed features effectively serve downstream congestion prediction tasks, the training of the compressed model should not only pursue high-fidelity reconstruction at the data level, but should focus on retaining information crucial to prediction. Core traffic state features (congestion index) directly determine the predicted value, while key road attributes (such as grade and historical patterns) provide important contextual priors. Therefore, a loss function is needed that can differentiate the reconstruction accuracy of these two types of information. This loss function ensures that the compressed model, while reducing dimensionality, effectively retains the core traffic state information and key static / quasi-static attributes crucial to downstream prediction tasks.

[0128] 5. Key parameter: Maximum length of the super-edge Node embedding dimension 20, node compression dimension Hyperedge compression dimension Number of message passing iterations .

[0129] 6. Input / output definition:

[0130] (1) Input: a single sub-signal from the training set / validation set / test set Fixed topology object Node dynamic attributes ;

[0131] in, It is a node timestamp The dynamic attribute vector below, with dimension 1 , is one of the inputs to the compression model.

[0132] (2) Output: Node compression features Hyper-edge compression features (For subsequent analysis and prediction), additional outputs during the training phase reconstruct features (for supervised training).

[0133] in, :node The initial compressed feature vector has a dimension of , is one of the outputs of the compression model. : Super-edge The deep compressed feature vector, with dimension , is one of the outputs of the compression model. Define the initial compression features of nodes in a 5-dimensional or 5-dimensional real space. The value space of is determined by the compression ratio. Decide. In a 3D real space, define the hyperedge depth compression feature. The value space of .

[0134] This invention proposes a SetMP-based topological feature compression method in step S3.2. Through a hierarchical strategy of "hyperedge construction - preliminary dimensionality reduction - deep compression," it achieves efficient compression of traffic data while preserving high-order topological relationships and core spatiotemporal features of the road network, thus reducing computational complexity. Furthermore, it minimizes the weighted MSE loss function between the original and reconstructed features. By rationally allocating the weights of core features (congestion index sequence) and key attributes (road grade weight, historical congestion intensity), it ensures that crucial information essential to the prediction task is not lost during the compression process.

[0135] S4: The node compression features are fused with the corresponding hyperedge compression features and input into the prediction model based on graph attention mechanism to predict the traffic congestion status of the target road segment at least one future time point.

[0136] Step S4 is used to implement intelligent traffic state analysis and prediction based on the GATv2 (Graph Attention Network v2) model. GATv2 is a graph neural network model. Utilizing low-dimensional features after topological compression, the GATv2 model captures the dynamic relationships between road segments. GATv2 relaxes the monotonicity constraint of attention weights in traditional graph attention networks, better adapting to the complex spatiotemporal relationships of traffic data, and achieving accurate traffic state analysis and prediction.

[0137] S4.1: Model Structure

[0138] 1. Graph Structure Definition: The adjacency relationships in the GATv2 model are constructed based on the physical connectivity of the road network—using nodes as the basis. Based on this, the physical connectivity between road segments (connection in the same direction, intersection connectivity) is represented by undirected edges. This forms an adjacency matrix, ensuring that the model captures the local correlation features of the road network.

[0139] in, : A set of road segment nodes, containing Each road segment node in Guiyang City is a basic node unit of the GATv2 graph structure. An undirected edge set, constructed based on the physical connectivity of the road network. Edges correspond to same-direction connections or intersection connections between road segments, used to generate the adjacency matrix of GATv2. Adjacency matrix: based on a set of nodes. and undirected edge set The constructed matrix represents the connection relationship between nodes in the GATv2 model, ensuring that the model captures the local correlation features of the road network.

[0140] 2. Model Depth: The prediction model has a total of 4 layers, such as... Figure 5 As shown, it includes an input layer, two attention layers, and an output layer, which ensures expressive power while avoiding overfitting.

[0141] (1) Input layer: compresses node features Compression feature of the unique hyperedge The features are concatenated (each node corresponds to only one hyperedge, with no dimensional conflicts) to form a fused feature, which serves as the input. The input dimension is... ( ).

[0142] in, :node The initial compressed feature vector has a dimension of (correspond )or (correspond ), is a component of the GATv2 input layer. :node Belongs to the only super edge The deep compressed feature vector, with dimension It is a component of the GATv2 input layer. The input dimensions of the GATv2 model, corresponding to The time value is The dimension is determined by the sum of the dimensions of the nodes and the hyperedge compression features.

[0143] (2) Attention layer: used to dynamically aggregate attention on fused features based on the graph structure constructed from the physical connection relationship of the road network. The attention layer consists of two GATv2 layers. The first layer has a hidden dimension of 64 (8 attention heads, capturing multi-dimensional associations), and the second layer has a hidden dimension of 32 (4 attention heads), with the activation function ReLU.

[0144] The GATv2 layer is the core layer of the graph attention network, employing a two-layer structure. The first layer has 64 hidden dimensions and 8 attention heads, while the second layer has 32 hidden dimensions and 4 attention heads, with ReLU activation function. The number of attention heads is a key parameter of the GATv2 layer, used to capture multi-dimensional relationships between nodes; the first layer has 8, and the second layer has 4. ReLU, the rectified linear unit activation function, is used after the output of the GATv2 attention layer to enhance the model's non-linear expressive power.

[0145] (3) Output layer: Configured to output the predicted traffic congestion index of the target road segment at multiple consecutive time points in the future. The output layer is a single fully connected layer (output dimension...). The congestion index is predicted for four future timestamps (5 minutes x 4 = 20 minutes for core road sections and 15 minutes x 4 = 1 hour for secondary roads).

[0146] The fully connected layer is the output layer of the GATv2 model, a single-layer structure with an output dimension of [missing information]. This corresponds to the congestion index prediction values ​​for the next four time stamps.

[0147] S4.2: Model Training

[0148] 1. Training data: Node / hyperedge compressed features of the training set The training set corresponds to congestion index labels for the next four timestamps; the validation set is used to monitor the training effect, and the test set is used for final evaluation.

[0149] 2. Input and Output Definitions:

[0150] Input: Current Node compression features at time step Compression feature of the unique hyperedge and fixed topology objects ;

[0151] Output: Future Traffic congestion index at any given time ( (Level), the corresponding relationship is: the time sampling interval for the core road section is 5 minutes, and 4 timestamps are... Prediction within minutes; branch time sampling interval is 15 minutes, 4 timestamps. Predictions within minutes (1 hour) are available to adapt to the traffic control response needs of different road sections.

[0152] : Current timestamp, representing the time node for traffic data collection, with a sampling interval of 5 minutes for core road sections and 15 minutes for branch roads. : A fixed topology object containing a collection of nodes and superedge set Its adjacency matrix is ​​constructed based on the physical connectivity of the road network and is one of the inputs to the GATv2 model. The four consecutive timestamps correspond to the predicted time range within 20 minutes for the core road section and within 1 hour for the branch road.

[0153] 3. Training configuration:

[0154] (1) Optimizer: Adam optimizer (learning rate) Numerical stability parameters (Default value), weight decay (used to suppress overfitting).

[0155] Where Adam is the adaptive moment estimator optimizer used for training the GATv2 prediction model. lr is the learning rate, a parameter of the Adam optimizer that controls the step size for updating model parameters; here, it is set to 0.001. Numerical stability parameter, a parameter of the Adam optimizer, used to avoid a denominator of 0; here it is set to a value of [value missing]. (Default value). Weight decay: The regularization parameter of the Adam optimizer, with a value of [value to be filled in]. This is used to suppress the risk of overfitting in the GATv2 model.

[0156] (2) Training rounds: 200 rounds, batch size 32.

[0157] (3) Training strategy: The training of the GATv2 model adopts a step-by-step training strategy. First, the parameters of the compressed model that have been trained in the SetMP mechanism are fixed; then, only the parameters of the prediction model are updated.

[0158] (4) Loss Function: The training objective of the GATv2 model is to minimize the mean squared error (MSE) between the predicted congestion index and the actual value at multiple consecutive future time points. Specifically, the MSE loss (mean squared error between the predicted congestion index and the actual value) is used, and the formula for the loss function is:

[0159]

[0160] The mean squared error loss function of the GATv2 prediction model is used to measure the difference between the predicted congestion index and the actual value. The total number of road segment nodes, included in a single batch of collected data. Each road segment node. The number of sub-signals contained in the training set, corresponding to the first... The total number of sub-signals in time series data. :node In the In the future of each sub-signal The actual congestion index (level 0-10) for each timestamp. Nodes predicted by the GATv2 model In the In the future of each sub-signal Congestion index (level 0-10) for each timestamp. : The square of the L2 norm of the vector, used here to calculate the difference between the true and predicted values ​​of a single node, a single sub-signal, and a single future timestamp. Sub-signal index, with a value range of 1 to... , corresponding to different sub-signals in the training set. : Constraints, representing nodes Belongs to the set of nodes . : Future timestamp index, with a value range of 1 to 4, corresponding to the future to These are the four timestamps.

[0161] In summary, the intelligent analysis method for traffic big data based on topology compression and feature dimensionality reduction provided by the above embodiments of the present invention has the following platform interface: Figure 6As shown, its core innovation lies in constructing a collaborative optimization framework of "data-structure-feature-prediction". First, it proposes a hyperedge construction method based on feature similarity clustering, upgrading the traditional road network map to a more semantically rich node-hyperedge two-layer topology, achieving efficient representation of complex traffic associations. Second, it designs a hierarchical compression mechanism (SetMP) for "node-hyperedge" sets, combining a two-stage strategy of "pre-training topology construction" and "fixed topology training compression", significantly reducing data dimensionality and computational complexity while ensuring no loss of core information. Finally, by combining the compressed topological features with the physically connected GATv2 prediction model and adopting a training strategy of "freezing the compression module first, then fine-tuning the prediction layer", it achieves high-precision and high-efficiency traffic state prediction. This method integrates the entire chain from data structuring and feature engineering to model training, effectively solving the problems of high computational load, feature redundancy, and insufficient prediction accuracy for large-scale traffic data.

[0162] By comparing the performance advantages of the algorithm of this invention with three existing algorithms, such as... Figure 7 As shown, this scheme (GATv2) reduces MAE by 41% compared to the traditional GCN model and by 34% compared to GraphSAGE. While maintaining a high compression ratio of 1 / 3, the number of parameters is controlled within a reasonable range to meet the needs of real-time prediction. The performance index of MAE 0.8 fully meets the requirements of traffic control decision-making (error level 1).

[0163] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0164] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A method for intelligent analysis of traffic big data based on topology compression and feature dimensionality reduction, characterized in that, Includes the following steps: S1: Obtain the time-series traffic status data of each road segment node in the target road network, and perform data cleaning and standardization processing. Divide the standardized dataset into training set, validation set and test set in time order. S2: Based on the training set data, calculate the correlation degree between road segment nodes, and aggregate multiple nodes into hyperedges according to the correlation degree to construct a road network topology structure containing a node layer and a hyperedge layer; wherein, each node uniquely belongs to a hyperedge; S3: Based on the road network topology, the features of the nodes are encoded and compressed using a set-based message passing mechanism to obtain the dimension-reduced compressed node features and hyperedge compressed features. S4: The node compression features are fused with the corresponding hyperedge compression features and input into the prediction model based on graph attention mechanism to predict the traffic congestion status of the target road segment at least one future time point.

2. The method according to claim 1, characterized in that, In step S2, a road network topology structure including a node layer and a hyperedge layer is constructed based on the training set data, specifically including: Set an upper limit p for the number of nodes that a single hyperedge can contain, and determine the total number of hyperedges K based on the total number of nodes N; Based solely on the node features of the training set, a pre-trained encoder maps each node feature into a node embedding vector; Based on the node embedding vectors, the similarity between nodes is calculated to form a similarity matrix that represents the degree of association between nodes; Using the similarity matrix as input, the K-Means clustering algorithm is used to divide all nodes into K disjoint sets, and each set is defined as a hyperedge. Based on the node set and hyperedge set obtained by the division, a fixed topology object is generated. The topology object remains unchanged after construction, and the data processing of the validation set and the test set reuses the topology object.

3. The method according to claim 2, characterized in that, When performing clustering, if the number of nodes in a generated hyperedge exceeds p, then clustering is performed again within that hyperedge to ensure that the number of nodes in each generated hyperedge does not exceed p.

4. The method according to claim 1, characterized in that, In step S2, constructing the road network topology also includes a topology attribute enhancement step, specifically including: attaching a dynamic attribute vector to each node in the topology, wherein the dynamic attribute vector includes road level information, historical congestion patterns, time factors, and spatial flow; and calculating a weight for each hyperedge in the topology, wherein the weight is determined based on the similarity between the nodes constituting the hyperedge.

5. The method according to claim 1, characterized in that, Step S3, the encoding and compression of node features using a collective message passing mechanism, includes the following steps: Topological feature extraction steps: Based on the road network topology, the traffic state temporal features and additional attribute information of the nodes are fused to generate node embedding features and hyperedge embedding features containing higher-order associations; Topological feature compression steps: Based on the node embedding features and hyperedge embedding features, preliminary dimensionality reduction and deep compression are performed through a set-style message passing system that includes a two-layer structure of nodes and hyperedges, resulting in node compression features and hyperedge compression features with reduced dimensionality.

6. The method according to claim 5, characterized in that, The topological feature extraction step specifically includes: Node embedding optimization: The core traffic state time-series features of nodes are concatenated with dynamic attribute vectors, and the node embedding features are updated through neural network mapping; Hyperedge embedding generation: For each hyperedge, perform permutation-invariant aggregation on the node embedding features of all nodes inside it, and then map the aggregation result through a neural network to generate hyperedge embedding features.

7. The method according to claim 5, characterized in that, The topology feature compression step specifically includes: Preliminary dimensionality reduction: For each node, its core features, node embedding features, and the hyperedge embedding features of its unique hyperedge are fused together, and a low-dimensional preliminary compressed node feature is generated by mapping through the first neural network. Deep compression: For each hyperedge, its hyperedge embedding features, the mean of the core features of all internal nodes, and the mean of the initial compressed features of the nodes are fused together, and the hyperedge compressed features are generated by mapping through a second neural network.

8. The method according to claim 2, characterized in that, The compression model corresponding to the aggregated message passing mechanism is obtained through training: in the first stage, the node embedding output by the encoder is used to assist in constructing the road network topology; in the second stage, the compression model is trained based on the fixed road network topology. Furthermore, the training objective of the compressed model is to reconstruct the core features and key road attributes of the original traffic state through the decoder, and the loss function used is a weighted loss function that assigns different weights to the reconstruction errors of the core features and the reconstruction errors of the attribute features.

9. The method according to claim 1, characterized in that, The prediction model based on the graph attention mechanism adopts the GATv2 model, and the structure of the prediction model includes: The input layer takes as input a fused feature formed by concatenating the node compression feature with the compression feature of the unique hyperedge to which it belongs; The attention layer, which includes two GATv2 layers, is used to dynamically aggregate the fused features based on a graph structure constructed from the physical connectivity of the road network. The output layer is configured to output the predicted traffic congestion index values ​​of the target road segment at multiple consecutive time points in the future.

10. The method according to claim 1, characterized in that, The training process of the prediction model includes: a) A step-by-step training strategy is adopted. First, the parameters of the compressed model that has been trained in the collective message passing mechanism are fixed; then, only the parameters of the prediction model are updated. b) The training objective of the prediction model is to minimize the mean square error between the predicted congestion index and the actual value for multiple consecutive future time points.

Images