Traffic flow prediction method based on multi-view embedding and adversarial networks

Abstract

This invention discloses a traffic flow prediction method based on multi-view embedding and adversarial networks, comprising: constructing an edge set and a node set of origin-destination pairs; dividing the edge set into multiple subsets at fixed time intervals; aggregating the origin and destination points of each subset to generate a sequence of real origin-destination demand matrices; acquiring continuous-time information, node memory information, and discrete-time information of nodes from both continuous-time and discrete-time perspectives, and then jointly updating them to generate node temporal embedding vectors; inputting the node temporal embedding vectors into a graph attention module, fusing neighbor information to obtain the final embedding vectors of the nodes; and reconstructing the final embedding vectors of the nodes using a Gaussian kernel function in the generator of the adversarial network to obtain the reconstructed origin-destination demand matrix. This invention can solve the problem of low prediction accuracy caused by data sparsity and insufficient information acquisition.

Description

Technical Field

[0001] This invention relates to the field of traffic data processing technology, and more specifically to a traffic flow prediction method based on multi-view embedding and adversarial networks. Background Technology

[0002] In recent years, with the accelerated development of the automotive industry and the upgrading of intelligent transportation systems, origin-destination traffic flow prediction has become a core technology for alleviating traffic congestion and optimizing road network resource allocation. Origin-destination traffic flow prediction primarily uses historical traffic data to predict traffic flow patterns between different areas in the next time period. This technology can help traffic management departments and travel service platforms accurately capture the spatiotemporal distribution patterns of origin-destination traffic from massive amounts of dynamic data, providing stronger scientific basis for real-time route planning, dynamic pricing, and emergency dispatch decisions, and is also of great significance to the development of transportation networks.

[0003] However, due to limitations in sampling conditions and increased privacy requirements, the actual traffic flow data collected is usually highly sparse. This sparsity increases the difficulty of mining hidden spatiotemporal patterns in traffic data, reduces robustness, and thus affects prediction accuracy. In addition, existing methods often take a single perspective, resulting in insufficient information acquisition, which also affects prediction accuracy. Summary of the Invention

[0004] In view of this, the present invention provides a traffic flow prediction method based on multi-view embedding and adversarial networks to solve the problem of low prediction accuracy caused by data sparsity and insufficient information acquisition.

[0005] A traffic flow prediction method based on multi-view embedding and adversarial networks includes:

[0006] Step S1: Read the trip record database, construct the edge set and node set of origin and destination pairs, then divide the edge set into multiple subsets according to a fixed time interval, aggregate the origin and destination points of each subset, and generate the real origin and destination demand matrix sequence.

[0007] Step S2: Based on the edge set and node set of the origin-end point pair and the actual origin-end point demand matrix sequence, obtain the continuous time information, node memory information and discrete time information of the node from both continuous time and discrete time perspectives, and then perform joint update to generate a node temporal embedding vector containing time information.

[0008] Step S3: Input the node temporal embedding vector into the graph attention module, fuse the neighbor information, and obtain the final embedding vector of the node;

[0009] In step S4, the generator of the adversarial network reconstructs the final embedding vector of the node using a Gaussian kernel function to obtain the reconstructed origin-destination demand matrix. The elements in the reconstructed origin-destination demand matrix represent the predicted traffic flow from the origin to the destination during the prediction time period. The discriminator of the adversarial network uses a multilayer perceptron to discriminate between the reconstructed origin-destination demand matrix and the real origin-destination demand matrix. The discriminant results are then combined with the sum of differences loss to construct a hybrid objective loss function.

[0010] The traffic flow prediction method based on multi-view embedding and adversarial networks provided by the present invention has the following beneficial effects:

[0011] (1) To address the problem that a single perspective is insufficient to effectively capture traffic flow changes, this invention obtains continuous time information, node memory information, and discrete time information of nodes from both continuous and discrete time perspectives, and then performs joint updates to generate node temporal embedding vectors containing time information. This solves the problem that existing technologies are limited to a single time perspective when capturing temporal information, resulting in insufficient information acquisition. The multi-perspective mechanism of this invention can simultaneously capture traffic fluctuations caused by short-term sudden events and long-term periodic patterns, capturing richer node time information to enhance the ability to capture traffic flow changes and thus improve prediction accuracy.

[0012] (2) To address the difficulty of mining hidden spatiotemporal patterns caused by data sparsity, this method adopts an adversarial network architecture, extracts the key structure of the network as noise input, generates diverse data through a generator to fill data gaps, reduces dependence on specific data, and improves generalization ability. The adversarial training loss function is obtained through a discriminator, and then combined with the difference summation loss function to construct a hybrid target loss function. This allows for multi-faceted evaluation of loss, improves robustness, and further enhances prediction accuracy. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the traffic flow prediction method based on multi-view embedding and adversarial networks provided in an embodiment of the present invention. Detailed Implementation

[0014] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain embodiments of the present invention, and should not be construed as limiting the present invention.

[0015] Please see Figure 1 The embodiments of the present invention provide a traffic flow prediction method based on multi-view embedding and adversarial networks, including steps S1-S4:

[0016] Step S1: Read the trip record database, construct the edge set and node set of origin-destination pair, then divide the edge set into multiple subsets according to a fixed time interval, aggregate the origin and destination points of each subset, and generate the real origin-destination demand matrix sequence.

[0017] In this embodiment, the dataset used in the trip record database is obtained from TLC Trip Record Data - TLC, and its main content is historical trip data of taxis in a certain city. The dataset provides the following main fields:

[0018] Supplier ID indicates the company or supplier providing this taxi service.

[0019] The timestamp when the passenger boards the taxi indicates the start time of the taxi trip.

[0020] The timestamp when the passenger gets off the taxi indicates the end time of the taxi trip.

[0021] Passenger count indicates the number of passengers taking this taxi trip.

[0022] Trip distance refers to the distance a taxi travels from its origin to its destination (usually in miles).

[0023] Rate code ID indicates the rate type for this trip, such as peak hours or nighttime rates.

[0024] Store and forward flags are used to indicate whether the record is later forwarded to the system due to a device malfunction.

[0025] The boarding point ID indicates the specific location where the passenger boarded the bus (usually based on a predefined geographical area or location code).

[0026] The drop-off point ID indicates the specific location where the passenger gets off (also based on a predefined geographical area or location code).

[0027] Payment method indicates how passengers will pay the taxi fare, such as cash, credit card, etc.

[0028] The basic fare represents the basic cost of this trip (excluding other expenses such as surcharges and taxes).

[0029] Additional fees refer to extra charges, such as baggage fees, night service fees, etc.

[0030] Total amount indicates the total cost that the passenger needs to pay.

[0031] Specifically, step S1 includes:

[0032] Step S11: By extracting travel events from the origin to the destination of the car from the trip record database, construct the edge set of origin-destination pairs. and node set Each origin-end pair is an edge with a timestamp;

[0033] Specifically, the origin and destination pairs The specific definition is This indicates that the car started from the starting point. Departure, at Time to reach the end , It is an edge feature.

[0034] Step S12: Set a fixed time interval According to a fixed time interval Set of edges The data is divided into multiple subsets. All origin-destination pairs within each subset are aggregated to generate a true origin-destination demand matrix. Multiple true origin-destination demand matrices constitute a true origin-destination demand matrix sequence. , , , They represent the 1st, 2nd, and 3rd respectively. The actual start and end point demand matrix for each time interval. elements in Represents a node To the node In time interval Domestic travel demand.

[0035] Step S2: Based on the edge set and node set of the origin-end point pair and the actual origin-end point demand matrix sequence, obtain the continuous time information, node memory information and discrete time information of the node from both continuous time and discrete time perspectives, and then perform joint update to generate a node temporal embedding vector containing time information.

[0036] Specifically, step S2 includes:

[0037] Step S21: From the perspective of continuous time, update the node information using the continuous time information in the original time data. First, update the edges in the edge set. Perform time encoding to obtain the time-encoded result of the edge. The encoding method is as follows:

[0038]

[0039] in, The destination node The node attribute features, where * is the multiplication symbol. It is a natural constant. It is a prediction of the start time. It is the time it takes for the car to travel from the starting point to the destination. It is the window size for the prediction time period;

[0040] After encoding, consecutive time information from the same edge is aggregated to obtain aggregated information. The calculation method is as follows:

[0041]

[0042] in, The destination node The latest memory representation vector, [] denotes the concatenation operation. It is an edge feature;

[0043] Divide all start-end pairs equally according to the number of edges. The batches are processed in parallel, and then the time information originating from the same node within each batch is aggregated according to the following formula:

[0044]

[0045] in, Indicates that the aggregation originates from the node. Information, It is a node To the node The edge, , Indicates the first One batch, Representing an edge Aggregated information, Indicates the preceding One dimension, It is a batch The starting node is The total number of edges, express And all subsequent dimensions, this calculation method will... The dimensions representing the memory of a node are averaged and aggregated. This average aggregation avoids memory fluctuations caused by high-frequency updates, maintaining the stability of the node representation. The dimensions thereafter represent the overall aggregation of time encoding and edge features, because edge features are additive, and the overall aggregation can reflect the total demand within the time window;

[0046] An aggregation operation is performed on each batch to update the node's historical memory. The aggregation information of all nodes is then concatenated to obtain the continuous aggregation information of all nodes in the current batch. At the same time, the information from the last batch aggregation is retained and used as continuous time information. For subsequent joint updates;

[0047] The aggregated continuous-time information is fused into the node memory, and the node memory information is updated through a gated recurrent network. The calculation method is as follows:

[0048]

[0049]

[0050]

[0051] in, This indicates the first update gate, which controls the proportion of historical information transmitted to the new time. This represents the first candidate hidden state, which is a combination of historical information and new input information; This represents the sigmoid function. Represents the hyperbolic tangent function. It is the first learnable parameter matrix. It is the second learnable parameter matrix. It is the first bias vector. It is the second bias vector. It is Hadamaji. It is the node memory information before the update. It is the updated node memory information;

[0052] Step S22: From a discrete-time perspective, the actual origin-destination demand matrix sequence is processed into a feature matrix sequence containing key information. The key structure is obtained by calculating the symmetric Laplace matrix of the actual origin-destination demand matrix. The calculation method is as follows:

[0053]

[0054] in, It is a matrix The symmetric Laplace matrix, It is the identity matrix. It is a matrix The degree matrix;

[0055] Then As noise input, it is concatenated with the node attribute matrix, and then dimensionality is reduced using Gaussian random projection to obtain the final feature matrix. The calculation method is as follows:

[0056]

[0057] in, It is a node attribute matrix. It is a random matrix whose elements are independently drawn from a standard Gaussian distribution;

[0058] The final feature matrix is ​​used to capture discrete-time information, and a gated recurrent network is used for memory updates. The calculation method is as follows:

[0059]

[0060]

[0061]

[0062] in, This indicates the second update gate, which controls the proportion of historical information transmitted to the new time. This represents the second candidate hidden state, which is a combination of historical information and new input information; It is the third learnable parameter matrix. It is the fourth learnable parameter matrix. It is the third bias vector. It is the fourth bias vector. It is the node memory of discrete-time information before the update. It is the node memory of the updated discrete-time information, which will Discrete-time information in subsequent joint updates ;

[0063] Step S23: Merge continuous-time information and discrete-time information into the memory node. The input information consists of three parts: , , The input quantities of continuous-time information and discrete-time information are controlled by two gates, thereby generating a node temporal embedding vector containing time information. The calculation method is as follows:

[0064]

[0065]

[0066]

[0067]

[0068]

[0069] in, It controls the amount of input information over continuous time. It controls the input amount of discrete-time information. , , , These are the fifth, sixth, seventh, and eighth learnable parameter matrices, respectively. , , , These are the fifth, sixth, seventh, and eighth bias vectors, respectively. This indicates the amount of information combined between node memory and continuous time information. This represents the amount of information combined between node memory and discrete-time information. This represents a matrix composed of node temporal embedding vectors containing time information.

[0070] Step S3: Input the node temporal embedding vector into the graph attention module, fuse the neighbor information, and obtain the final embedding vector of the node.

[0071] Specifically, step S3 includes:

[0072] Step S31, for node Generate supplementary time information The calculation method is as follows:

[0073]

[0074] in, This represents the ninth learnable parameter matrix. It is a node The last time it was updated;

[0075] Step S32: Sample the neighbors of the node. Based on the node's temporal embedding vector, use multiple graph attention layers in the graph attention module to aggregate the information of the node and its neighbors. The inputs to the graph attention layers are the node's embedding representation and the set of neighboring nodes, respectively. The calculation method for the node's embedding representation is as follows:

[0076]

[0077]

[0078] in, It is the node corresponding to the first graph attention layer. Embedded representation, It is the first The nodes corresponding to the attention layer of the graph Embedded representation, yes Middle node Temporal embedding vectors, It was before Nodes generated by the graph attention layer Embedded representation, Indicates a serial operation;

[0079] Step S33, for nodes and the neighboring nodes Calculate attention weight coefficients The calculation method is as follows:

[0080]

[0081]

[0082] in, It is a node and nodes Attention weight value, It is the parameter vector of the attention mechanism. Indicates transpose. It is a learnable parameter matrix. It is a node Embedded representation, It is a normalization function. It is a node The set of neighboring nodes;

[0083] Step S34: The attention weight coefficients are weighted and summed with the node embedding vectors to obtain the node embedding vector that aggregates neighbor information. The calculation method is as follows:

[0084]

[0085] in, Represents a nonlinear function. It is the tenth learnable parameter matrix;

[0086] Nodes are generated through iterative weighted calculations across multiple graph attention layers. The final embedding vector .

[0087] In step S4, the generator of the adversarial network reconstructs the final embedding vector of the node using a Gaussian kernel function to obtain the reconstructed origin-destination demand matrix. The elements in the reconstructed origin-destination demand matrix represent the predicted traffic flow from the origin to the destination during the prediction time period. The discriminator of the adversarial network uses a multilayer perceptron to discriminate between the reconstructed origin-destination demand matrix and the real origin-destination demand matrix. The discriminant results are then combined with the sum of differences loss to construct a hybrid objective loss function.

[0088] The generator employs a reconstruction method based on Gaussian kernel functions to generate the start-end point demand matrix for reconstruction. ,matrix elements in Indicates the time period during which prediction is made. From the starting point To the finish line The traffic flow forecast is calculated as follows:

[0089]

[0090]

[0091]

[0092] in, It is the auxiliary mapping representation obtained by concatenating the final embedding vector of the node with the node attribute matrix; It is a multilayer perceptron, whose output layer uses the tanh function as the activation function. It is the concatenation result of the final embedding vectors of all nodes. It is an adaptive parameter matrix. yes The Middle Line number Column elements, It is a node The final embedding vector, This represents the square of the L2 norm.

[0093] During the training phase, the reconstructed origin-end point demand matrix As training samples. During the prediction phase, the reconstructed origin-end point demand matrix. It is used directly as the prediction result.

[0094] In this embodiment, the core architecture of the discriminator adopts a multilayer perceptron, and uses an alternating sampling mechanism to receive the reconstructed origin-destination demand matrix. and the actual origin and destination demand matrix By passing transformations layer by layer, the multilayer perceptron gradually extracts high-level features from the input data. Finally, the output layer uses the sigmoid function, and the output value represents the probability vector that the input data belongs to the true data distribution.

[0095] Specifically, hybrid objective loss function The expression is:

[0096]

[0097]

[0098]

[0099] in, This is the sum of differences loss function, used to directly represent the sum of differences between the predicted generated value and the true value; To counteract the training loss function, and They are respectively and The weighting coefficients, The total number of nodes. It is a real origin-destination demand matrix. The elements in The discriminant output matrix represents the first... The probability value that a row element belongs to the real data.

[0100] In summary, the traffic flow prediction method based on multi-view embedding and adversarial networks according to the above embodiments has the following beneficial effects:

[0101] (1) To address the problem that a single perspective is insufficient to effectively capture traffic flow changes, this invention obtains continuous time information, node memory information, and discrete time information of nodes from both continuous and discrete time perspectives, and then performs joint updates to generate node temporal embedding vectors containing time information. This solves the problem that existing technologies are limited to a single time perspective when capturing temporal information, resulting in insufficient information acquisition. The multi-perspective mechanism of this invention can simultaneously capture traffic fluctuations caused by short-term sudden events and long-term periodic patterns, capturing richer node time information to enhance the ability to capture traffic flow changes and thus improve prediction accuracy.

[0102] (2) To address the difficulty of mining hidden spatiotemporal patterns caused by data sparsity, this method adopts an adversarial network architecture, extracts the key structure of the network as noise input, generates diverse data through a generator to fill data gaps, reduces dependence on specific data, and improves generalization ability. The adversarial training loss function is obtained through a discriminator, and then combined with the difference summation loss function to construct a hybrid target loss function. This allows for multi-faceted evaluation of loss, improves robustness, and further enhances prediction accuracy.

[0103] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1.A traffic flow prediction method based on multi-view embedding and adversarial network, characterized in that, Comprise: Step S1, read the trip record database, build the edge set and node set of the origin-destination pair, and then divide the edge set into multiple subsets according to the fixed time interval, aggregate the origin-destination pairs of each subset, and generate the real origin-destination demand matrix sequence; Step S2, according to the edge set and node set of the origin-destination pair and the real origin-destination demand matrix sequence, respectively from the continuous time and discrete time two perspectives, obtain the continuous time information of the node, the node memory information and the discrete time information, and then jointly update to generate the node time sequence embedding vector containing time information; Step S3, input the node time sequence embedding vector into the graph attention module, fuse the neighbor information, and obtain the final embedding vector of the node; Step S4, the generator of the adversarial network uses the Gaussian kernel function to reconstruct the final embedding vector of the node, obtains the reconstructed origin-destination demand matrix, the elements in the reconstructed origin-destination demand matrix represent the traffic flow prediction value from the origin to the destination in the prediction time period, the discriminator of the adversarial network uses the multilayer perception to discriminate and train the reconstructed origin-destination demand matrix and the real origin-destination demand matrix, and then combines the discrimination result and the difference sum loss to build a hybrid objective loss function; Wherein, step S1 comprises: Step S11, constructing a set of edges of origin-destination pairs by extracting the travel events of the car from the origin to the destination from the trip record database and a set of nodes Each origin-destination pair is a time-stamped edge; Step S12, setting a fixed time interval , according to a fixed time interval The edge set is divided into multiple subsets, and all the origin-destination pairs inside the subsets are aggregated to generate a real origin-destination demand matrix, and multiple real origin-destination demand matrices constitute a real origin-destination demand matrix sequence , , , respectively represent the first, second, and real origin-destination demand matrix of the first, second, and Wherein, step S2 comprises: Step S21, from the perspective of continuous time, the node information is updated by the continuous time information in the original time data. First, the edges in the edge set are time-coded to obtain the time-coded results of the edges. After coding, the continuous time information from the same edge is aggregated to obtain aggregation information. All the origin-destination pairs are evenly divided into batches according to the number of edges, and then the time information from the same node within the batch is aggregated. aggregated.​ The aggregation operation is performed on each batch for updating the node historical memory, and the aggregated information of all nodes is spliced to obtain continuous aggregated information of all nodes in the current batch , while retaining the information aggregated in the last batch, taking the information aggregated in the last batch as continuous time information for subsequent joint update; Fusing the aggregated continuous time information into the node memory, updating the node memory information through a gated recurrent network to obtain the node memory information before updating and the node memory information after updating ; Step S22, from the perspective of discrete time, the real origin-destination demand matrix sequence is processed into a feature matrix sequence containing key information, the key structure is obtained by calculating the symmetric Laplacian matrix of the real origin-destination demand matrix; then the symmetric Laplacian matrix is taken as noise input, spliced with the node attribute matrix, dimensionality reduction is carried out through Gaussian random projection to obtain the final feature matrix; the final feature matrix is used to capture the discrete time information, memory update is carried out using the gated recurrent network to obtain the discrete time information ​ Step S23, fuse the continuous time information and the discrete time information into the memory node, the input information has three, respectively , , , the input amount of the continuous time information and the discrete time information is controlled through two gates, thereby generating a node time sequence embedding vector containing time information. 2.The traffic flow prediction method based on multi-view embedding and adversarial network according to claim 1, characterized in that, Step S2 specifically comprises: Step S21, firstly, time-encodes the edges in the edge set to obtain time-encoding results of the edges The encoding mode is as follows:​ wherein, is a node attribute feature of the destination node , is a multiplication symbol, is a natural constant, is a time of prediction start, is a time of the car from the start point to the end point, is a window size of the prediction time period; After encoding, the continuous time information from the same edge is aggregated to obtain aggregated information The calculation method is as follows: wherein, is the latest memory representation vector of the destination node , [] denotes the concatenation operation, is the edge feature; All pairs of origin-destination points are divided equally by the number of edges into batches that are processed in parallel, and then the time information from the same node within a batch is aggregated according to the following formula: wherein, represents aggregated information from nodes , is a node to a node , , represents the th batch, represents aggregated information of edges , represents the first dimensions, is the total number of edges with the starting node in the batch , represents and all dimensions after it; The aggregation operation is performed on each batch for updating the node historical memory, and the aggregation information of all nodes is spliced to obtain continuous aggregation information of all nodes in the current batch , while retaining the information aggregated in the last batch, taking the information aggregated in the last batch as continuous time information for subsequent joint update; Fuse the aggregated continuous time information into the node memory, update the node memory information through the gated recurrent network, and the calculation method is as follows: wherein, denotes a first update gate, denotes a first candidate hidden state, denotes a sigmoid function, tanh denotes a hyperbolic tangent function, is a first learnable parameter matrix, is a second learnable parameter matrix, is a first bias vector, is a second bias vector, is a Hadamard product, is a node memory information before update, is a node memory information after update; Step S22, from the perspective of discrete time, process the real origin-destination demand matrix sequence into a feature matrix sequence containing key information, obtain the key structure by calculating the symmetric Laplacian matrix of the real origin-destination demand matrix, and the calculation method is as follows: wherein is a symmetric Laplacian matrix of the matrix , is an identity matrix, is a degree matrix of the matrix ​ Then the As noise input, concatenate with the node attribute matrix, and get the final feature matrix by dimensionality reduction through Gaussian random projection The calculation method is as follows: wherein, is a node attribute matrix, is a random matrix; Capture the discrete time information using the final feature matrix, and update the memory using the gated recurrent network, and the calculation method is as follows: wherein, denotes a second update gate, denotes a second candidate hidden state, is a third learnable parameter matrix, is a fourth learnable parameter matrix, is a third bias vector, is a fourth bias vector, is a node memory of discrete-time information before update, is a node memory of discrete-time information after update, and is used as discrete-time information in a subsequent joint update ; Step S23, fuse the continuous time information and the discrete time information into the memory node, the input information has three, respectively , , , the input amount of continuous time information and discrete time information is controlled by two gates, so as to generate the node time sequence embedding vector containing time information, and the calculation method is as follows: wherein, is an input quantity of control continuous-time information, is an input quantity of control discrete-time information, , , , are the fifth, sixth, seventh, and eighth learnable parameter matrices, respectively, , , , are the fifth, sixth, seventh, and eighth bias vectors, respectively, denotes a combination quantity of node memory information and continuous-time information, denotes a combination quantity of node memory information and discrete-time information, denotes a matrix composed of node time sequence embedding vectors containing time information. 3.The traffic flow prediction method based on multi-view embedding and adversarial network according to claim 2, characterized in that, Step S3 specifically comprises: Step S31, for the node generating a complementary time information , the calculation method is as follows: wherein, denotes a ninth learnable parameter matrix, is a node time of last update; Step S32, sample the neighbors of the node, based on the node time sequence embedding vector, use multiple graph attention layers in the graph attention module to aggregate the information of the node and its neighbors, the input of the graph attention layer is the embedding representation of the node and the neighbor node set, and the calculation method of the embedding representation of the node is as follows: wherein, is the embedding representation of the node corresponding to the first graph attention layer , is the embedding representation of the node corresponding to the first graph attention layer , is the embedding representation of the node corresponding to the first graph attention layer , is the time series embedding vector of the node in the graph G, is the embedding representation of the node generated by the first graph attention layer, is the embedding representation of the node generated by the first graph attention layer, represents a concatenation operation; Step S33, the node and the first neighbor node calculates the attention weight coefficient , and the calculation method is as follows: wherein, is a node and an attention weight value of a node , is a parameter vector of an attention mechanism, denotes a transpose, is a learnable parameter matrix, is an embedding representation of a node , is a normalization function, is a set of neighbor nodes of a node ; Step S34, the attention weight coefficient is weighted and summed with the embedding vector of the node to obtain a node embedding vector of aggregated neighbor information The calculation method is as follows: wherein, denotes a non-linear function, is a tenth learnable parameter matrix; After iterative weighted computation through multiple graph attention layers, a final embedding vector of a node is generated .​ 4.The traffic flow prediction method based on multi-view embedding and adversarial network according to claim 3, characterized in that, In step S4, the generator generates a reconstructed origin-destination demand matrix using a reconstruction method based on a Gaussian kernel function , where the element in the matrix represents the predicted traffic flow value from the origin to the destination in the prediction time period , and the calculation method is as follows: wherein, is an auxiliary mapping representation obtained by concatenating the final embedding vectors of the nodes with the node attribute matrix, is a multi-layer perceptron, is a concatenation result of the final embedding vectors of all nodes, is an adaptive parameter matrix, is the element in the i-th row and the j-th column of the matrix, the element in the i-th row and the j-th column of the matrix, the element in the i-th row and the j-th column of the matrix, the final embedding vector of the node, the final embedding vector of the node, denotes the square of the L2 norm. 5.The traffic flow prediction method based on multi-view embedding and adversarial network according to claim 4, characterized in that, In step S4, the target loss function is mixed The expression is: wherein, is a difference sum loss function, is an adversarial training loss function, and are weight coefficients of and respectively, is the total number of nodes, is an element in the real origin-destination demand matrix is an element in the real origin-destination demand matrix represents the probability value that the element in the row of the discriminator output matrix belongs to the real data.

Images