Traffic Resource Allocation Method, System and Medium Based on Confidence Map Convolutional Network

Abstract

The present invention discloses a traffic resource allocation method, system and medium based on a confidence graph convolutional network. The traffic resource allocation method includes modeling the current actual traffic network and traffic stations; extracting features from the traffic network graph data generated by the modeling through a graph embedding method to determine a feature extraction vector; processing the feature extraction vector based on a deep belief network and propagating and updating it through a graph convolutional neural network, so as to update the feature representation of the nodes and change the label type of each node to generate a target topological structure, and then automatically allocate traffic resources to the corresponding traffic stations according to the target topological structure to achieve automatic allocation of traffic resources at one time. The present invention can respond in a timely manner to the resource allocation problem of traffic stations, thereby achieving a more optimal dynamic allocation of traffic resources.

Description

Technical Field

[0001] The present invention relates to the field of traffic resource allocation, and particularly to a traffic resource allocation method, system and medium based on a confidence graph convolutional network. Background Art

[0002] Broadly speaking, traffic resources refer to the basic designs and their combinations of all transportation modes under various technical conditions in the transportation industry. The optimal allocation of traffic resources is directly considered as the optimal combination of the infrastructure of various transportation modes. The intelligent traffic resource allocation system is first based on real sampling and data analysis. By abstracting and modeling the real scenario, and then using advanced analysis algorithms based on the modeling results to analyze the optimal allocation method of traffic resources.

[0003] In recent years, at the macroscopic level, China's transportation industry has achieved leapfrog development. The transportation network mainly composed of public road traffic and high-speed rail intercity traffic has spread widely across the country. Transportation hubs have been established in major cities, and cross-provincial transportation modes have gradually become rich. Both the passenger volume and the freight volume have achieved consecutive year-on-year growth for many years. At the microscopic level, each city has its own independent local traffic resources, such as shared bicycles, urban rails and buses, etc. However, with the expansion of the scale of the traffic resource network, the network topology has become more and more complex, which means that how to efficiently allocate traffic resources has become a difficult problem. Traditional traffic resource allocation cannot dynamically capture the changes in the traffic network, so it lacks the ability of dynamic adaptive allocation.

[0004] Traditional convolutional neural networks can only process data in Euclidean space, such as images, texts, voices, etc. The data in these fields has translational invariance. Translational invariance means that with any pixel point as the center, the same-sized local structure can be obtained. Translational invariance enables the definition of a globally shared convolutional kernel in the input data space, thereby defining a convolutional network. However, graph data is a kind of non-Euclidean space data, and the local structures of its each node are different, which makes translational invariance no longer satisfied, that is, graph data does not have translational invariance. Therefore, the basic operators (convolution and pooling) in traditional convolutional neural networks cannot achieve the application functions on graph data.

[0005] Therefore, there is a need for a traffic resource allocation method, system and medium based on a confidence graph convolutional network that can solve the above problems. Summary of the Invention

[0006] The purpose of the present invention is to provide a traffic resource allocation method, system and medium based on a confidence graph convolutional network for coping with the problem of dynamic resource allocation in a traffic network.

[0007] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] A traffic resource allocation method based on a confidence graph convolutional neural network. The traffic resource allocation method is used for the automatic allocation of traffic resources, and the traffic resource allocation method includes:

[0009] Model the current actual traffic network and traffic stations. The process includes converting the actual traffic network to generate an edge set of traffic network graph data, and converting some traffic stations to generate a node set of traffic network graph data;

[0010] Extract features from the traffic network graph data generated by modeling through a graph embedding method to determine a feature extraction vector. The feature extraction process includes obtaining the original feature representation of the node and inputting the original feature representation into the Skipgram model for processing to obtain the final feature extraction vector;

[0011] Process the feature extraction vector based on a deep belief network and propagate and update it through a graph convolutional neural network, thereby updating the feature representation of the node and changing the label type of each node to generate a target topological structure. Then, automatically allocate traffic resources to the corresponding traffic stations according to the target topological structure to achieve the automatic allocation of traffic resources once. The processing process based on the deep belief network includes unsupervised pre-training based on RBM and weight initialization through the CD-k method, thereby calculating the probability distribution of the hidden layer confidence neurons in the deep belief network being activated, and then updating the weights based on the probability distribution. The process of propagating and updating through the graph convolutional neural network includes aggregating the node set of the traffic network graph data through pooling propagation. The label type of the node represents the current traffic resource tension degree of the traffic station corresponding to the node. When the tension degree is higher, more traffic resources are allocated to the corresponding traffic station.

[0012] Further, the traffic resource allocation method further includes:

[0013] After completing the automatic allocation of traffic resources once, update the resource allocation situation of the current actual traffic network, and based on the updated actual traffic network data, perform the automatic allocation of traffic resources again. Among them, the method for updating the resource allocation situation of the current actual traffic network includes:

[0014] Update the feature matrix of all traffic nodes according to the change situation of traffic resources at each traffic station in the actual traffic network after the automatic allocation to obtain new traffic network graph data;

[0015] Use the GCN network model to perform Fourier transform on the new traffic network graph data and propagation between the node and its neighbor nodes. Among them, the features of the target nodes related to the target topological structure are aggregated according to the features of their neighbor nodes;

[0016] Update the label type of the target node according to the aggregation result, and at the same time start backpropagation to update the label types of the neighbor nodes of the target node.

[0017] Further, the method for obtaining the original feature representation of the node includes:

[0018] Repeat the following steps to collect multiple groups of node sequences: start from any node in the graph data and transfer between the nodes with equal probability.

[0019] Further, the method for inputting the original feature representation into the Skipgram model to process and obtain the final feature extraction vector includes:

[0020] Update some weights of the hidden layer during backpropagation, and determine whether to update the weights according to the node occurrence probability related to the weights, continue to sample some nodes, and control the probability of the nodes being sampled;

[0021] Among them, the probability of updating the weight is determined according to the following formula:

[0022]

[0023] In the formula, P(v) vi is the occurrence probability of node v i Z(v i ) is the weight of node v i Z(v j ) is the weight of node v j v i and v j come from the node set of the traffic network graph data;

[0024] Among them, the node occurrence probability is determined according to the following formula:

[0025]

[0026] In the formula, P(o i ) is the occurrence probability of node o i in the network constructed according to the existing weights, and Z(o i ) is the weight of node o i in the network constructed according to the existing weights.

[0027] Further, calculate the probability distribution of the hidden layer confidence neurons being activated in the deep belief network according to the following formula:

[0028]

[0029] In the formula, is a probability distribution, v (0) is a set of nodes from traffic network graph data, θ is the learning rate, W j is a vector matrix to be processed, b j is the bias.

[0030] Furthermore, the processing process based on the deep belief network further includes:

[0031] Continue Gibbs sampling (GIBBS sampling) according to the calculated probability distribution, and extract corresponding values for the neurons in the visible layer of the deep belief network for sampling reconstruction, so as to update and obtain new weights and biases. The process is represented by the following formula:

[0032] W←W+λ[x(h (0) =1)|z (0) z (0)T -x(x (1) =1|v(1))z (1)T

[0033] In the formula, W is the new weight matrix, λ and T are learning parameters, x is the sampling probability, h (0) is the hidden layer representation of the 0th layer. The calculation method of the said h (0) is that after the aggregation operation of the 0th layer, the results obtained by the nodes on different subgraphs are added together. z (0) is the output at the node level of the 0th layer, x (1) is the sampling probability of the 1st layer, z (1) is the output at the node level of the 1st layer, and v(1) is the set of nodes of the 1st layer.

[0034] Furthermore, aggregate the set of nodes of the traffic network graph data according to the following formula:

[0035]

[0036] In the formula, pool is the pooling operation, s is the vector of the s-th layer, θ is the learning rate, W is the adjacency vector, is the prediction result of node v i , k represents the k-th layer of the neural network, b is the offset, and v i comes from the set of nodes V of the traffic network graph data.

[0037] Furthermore, the process of updating based on the graph convolutional neural network further includes:

[0038] Update the feature representation of the node by performing gradient descent on the loss function, where the loss function is determined by the following formula:

[0039]

[0040] In the formula, L g (u z ) is the loss function, θ is the learning rate, and u z 、u v are the feature vectors of the traffic network graph data, T is the learning parameter, is the probability distribution.

[0041] A traffic resource allocation system based on a confidence graph convolutional neural network, the traffic resource allocation system includes:

[0042] A modeling unit configured to model the current actual traffic network and traffic stations to generate traffic network graph data;

[0043] A feature extraction unit configured to extract features from the traffic network graph data generated by modeling through a graph embedding method to determine a feature extraction vector;

[0044] A data processing unit configured to process the feature extraction vector based on a deep belief network and propagate and update it through a graph convolutional neural network, thereby updating the feature representation of the node and changing the label type of each node to generate a target topological structure, and then automatically allocating traffic resources to corresponding traffic stations according to the target topological structure to achieve automatic allocation of traffic resources at one time.

[0045] A storage medium configured to store a computer program, and the computer program is configured to execute the traffic resource allocation method described above.

[0046] The advantages of the present invention are: taking graph data as input, based on high-precision modeling of the real traffic network environment, using a confidence graph convolutional neural network that combines a deep belief convolutional neural network and a graph convolutional neural network to achieve automatic allocation of traffic resources, with high allocation efficiency, and can respond to the resource allocation problem of traffic stations in a timely manner, thereby achieving more optimal dynamic allocation of traffic resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for description in the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0048] Figure 1 It is a schematic diagram of the traffic resource allocation system provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0049] To enable those skilled in the art to better understand the solution of the present invention, and to more clearly understand the purpose, technical solution and its advantages of the present invention, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be noted that the implementation manners not depicted or described in the drawings are forms known to those of ordinary skill in the art. Additionally, although this document may provide examples containing parameters with specific values, it should be understood that the parameters need not exactly equal the corresponding values, but may approximate the corresponding values within an acceptable error tolerance or design constraint. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention. In addition, the terms "comprising" and "having" and any variations thereof in the specification and claims of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product or device that comprises a series of steps or units need not be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0050] In one embodiment of the present invention, a traffic resource allocation method based on a confidence graph convolutional neural network is provided. This traffic resource allocation method is used for the automated allocation of traffic resources.

[0051] First, model the current actual traffic network and traffic stations. In this embodiment, the actual traffic network is converted to generate an edge set of traffic network graph data, and part of the traffic stations are converted to generate a node set of traffic network graph data;

[0052] Secondly, use the graph embedding method to extract features from the traffic network graph data generated by modeling, so as to determine the feature extraction vector. In this embodiment, the process of feature extraction includes obtaining the original feature representation of the nodes. The obtaining method is to repeatedly perform the step of starting from any node in the graph data and transferring between nodes with equal probability, so as to collect multiple groups of node sequences, and then input the original feature representation, that is, the collected node sequences, into the Skipgram model for processing to obtain the final feature extraction vector. The specific processing process includes updating part of the weights of the hidden layer during backpropagation, and determining whether to update the weights according to the node appearance probability related to the weights, and continuing to sample some nodes, and controlling the sampling probability of the nodes.

[0053] Then, the feature extraction vectors are processed based on a deep belief network and propagated and updated through a graph convolutional neural network, so as to update the feature representation of the nodes and change the label type of each node, thereby generating a target topological structure, and then automatically allocating traffic resources to corresponding traffic stations according to the target topological structure to achieve the automatic allocation of traffic resources once. In this embodiment, the processing process based on the deep belief network includes unsupervised pre-training based on RBM and weight initialization by the CD-k method, so as to calculate the probability distribution of the activation of the hidden layer confidence neurons in the deep belief network, and then continue Gibbs sampling (GIBBS sampling) according to the calculated probability distribution, and extract corresponding values for the neurons in the visible layer of the deep belief network for sampling reconstruction, so as to update and obtain new weights and biases; the process of propagation and update based on the graph convolutional neural network includes aggregating the node set of the traffic network graph data through pooling propagation, where the update process further includes updating the feature representation of the nodes by performing gradient descent on the loss function.

[0054] It should be noted that the label type of a node represents the current traffic resource tension degree of the traffic station corresponding to the node. When the tension degree is higher, more traffic resources are allocated to the corresponding traffic station. In this embodiment, the label types of the nodes are set to include 1, 2, and 3, where 1 represents resource shortage, 2 represents moderate resources, and 3 represents sufficient resources. If the label type of a certain traffic station is 1, it means that traffic resources need to be invested (or more) in this traffic station; if the traffic station label is 3, it means that traffic resources do not need to be invested (or reduced) in this traffic station. The specific content of the label type can be set according to the actual application, and the protection scope of the present invention is not limited by the above 1, 2, and 3.

[0055] Finally, after completing the automatic allocation of traffic resources once, update the resource allocation situation of the current actual traffic network, and perform the automatic allocation of traffic resources again based on the updated actual traffic network data; among them, the method for updating the resource allocation situation of the current actual traffic network includes:

[0056] Update the feature matrix of all traffic nodes according to the change situation of traffic resources at each traffic station in the actual traffic network after automatic allocation to obtain new traffic network graph data;

[0057] Perform Fourier transform on the new traffic network graph data by using the GCN network model and propagation between nodes and neighbor nodes, where the features of the target nodes related to the target topological structure are aggregated according to the features of their neighbor nodes;

[0058] Update the label type of the target node according to the aggregation result, and at the same time start backpropagation to update the label types of the neighbor nodes of the target node. Loop the above process to achieve continuous dynamic and automated allocation of traffic resources at traffic stations, and timely solve the problems of resource shortage or overflow at stations.

[0059] In one embodiment of the present invention, the traffic resource allocation method includes the following steps:

[0060] (1) Realistic modeling

[0061] First, model the traffic network and important nodes in reality. According to the structural characteristics of the graph network, the traffic network will be abstracted into an edge set E{e1, e2, e3,..., e n}, where E represents the total set of edges, e1, e2, e3,..., e n represents n edges, e represents any one edge, and important traffic nodes will be abstracted into node set V{v1, v2, v3... v n}, where V represents the total set of nodes, v1, v2,..., v n represents n nodes, and v represents any one node. According to the interconnection relationship between networks, there will be interconnected edges between nodes, and at this time, the adjacency matrix W is used to describe the relationship.

[0062] This step can obtain the network data of the traffic resource network in reality and obtain a point set, an edge set, and an adjacency matrix to describe the traffic resource network.

[0063] (2) Feature extraction

[0064] This traffic resource allocation method uses graph embedding to extract features from the graph data generated by the modeling result.

[0065] First, use the ANRL model to obtain the original feature representation. The specific operation is to start from any point in the graph and transfer between nodes with equal probability, repeat the above steps a finite number of times, and finally collect a series of node sequences. After collecting enough sequences, use a method similar to processing text to input these sequences into the Skipgram model and finally obtain the embedding vector.

[0066] This traffic resource allocation method introduces negative sampling to accelerate the calculation speed to improve the performance of the homogeneity and structural equivalence of the embedding vector. Specifically, during backpropagation, some weights of the hidden layer are updated to reduce the calculation amount. At the same time, whether the weight is updated depends on the probability of the node related to the weight appearing. The greater the probability of the node appearing, the greater the probability of the weight related to it being updated. The probability of the weight being updated is determined according to the following formula:

[0067]

[0068] Wherein, P(v) vi is the occurrence probability of node v i , Z(v i ) is the weight of node v i , Z(v j ) is the weight of node v j , and v i , v j are from the node set of the traffic network diagram data;

[0069] On this basis, continue to sample some nodes in the network to save sampling time. In partial sampling, it is necessary to control the probability of nodes being sampled. The higher the frequency of node appearance, the smaller the probability of using it as the starting point for sampling, that is, the nodes that have been repeatedly sampled do not need to be resampled. Among them, the occurrence probability of nodes is determined according to the following formula:

[0070]

[0071] Wherein, P(o i ) is the occurrence probability of node o i in the network constructed according to the existing weights, and Z(o i ) is the weight of node o i in the network constructed according to the existing weights. Thus, the feature extraction vector W is finally obtained as the extraction result.

[0072] This step is to perform dimensionality reduction extraction on the features in the graph network generated in step (1), and prepare for obtaining the network data representation matrix in step (3) described below. The feature extraction tool used in this step is the ANRL model. The ANRL model respectively models node attributes and relationships through a dual-objective network. The finally obtained graph embedding vector is restricted by the attribute and relationship training data, and a graph embedding vector that fuses attribute and relationship information is obtained.

[0073] (3) Obtain the data representation of the graph network

[0074] Input the feature extraction result into the vector result W2 obtained by processing with a deep belief network and perform Fourier transform on it.

[0075] First, perform unsupervised pre-training based on RBM, and use the CD-k method for weight initialization:

[0076]

[0077] Wherein, W is the weight vector, a is the bias vector of the visible layer, and b is the bias vector of the hidden layer.

[0078] Then assign W to the visible layer and calculate the probability that the hidden layer confidence neurons are activated according to the following formula:

[0079]

[0080] In the formula, is the probability distribution, v (0) is the set of nodes from the traffic network graph data, θ is the learning rate, W j is the vector matrix to be processed, b j is the bias.

[0081] Then continue with Gibbs sampling (GIBBS sampling) based on the calculated probability distribution to sample and reconstruct the corresponding values of the neurons in the visible layer, and update to obtain new weights and biases:

[0082] W←W+λ[x(h (0) =1)|z (0) z (0)T -x(x (1) =1|v(1))z (1)T

[0083] In the formula, W is the new weight matrix, λ and T are learning parameters, x is the sampling probability, h (0) is the hidden layer representation of the 0th layer, and the calculation method of h (0) is that after the aggregation operation of the 0th layer, the results obtained by the nodes on different subgraphs are added together. z (0) is the output at the node level of the 0th layer, x (1) is the sampling probability of the 1st layer, z (1) is the output at the node level of the 1st layer, and v(1) is the set of nodes of the 1st layer.

[0084] The processing model used in this step is a deep belief network. As shown in Figure 1 , this deep belief network is composed of several neurons, and the component is the RBM restricted Boltzmann machine. The structure of the deep belief network is restricted to two layers, the visible layer and the hidden layer. There are connections between layers, but there are no connections between units within a layer. The hidden layer units are trained to capture the correlations of high-order data manifested in the visible layer.

[0085] (IV) Propagation

[0086] In this stage, propagate the target nodes, specifically through a propagation function. The propagation function can be average propagation or pooling propagation. In this embodiment, pooling propagation is used to continuously aggregate the data:

[0087]

[0088] where pool is the pooling operation, s is the vector of the s-th layer, θ is the learning rate, W is the adjacent vector, is the prediction result of node v i , k represents the k-th layer of the neural network, b is the offset, and v i comes from the node set V of the traffic network graph data.

[0089] Specifically, first perform a non-linear transformation on the adjacent node representation vectors of the target vertex, then perform a pooling operation (max pooling or average pooling), concatenate the obtained result with the representation vector of the target vertex, and finally perform another non-linear transformation to obtain the k-th layer representation vector of the target vertex.

[0090] (V) Update

[0091] In this stage, the update of node features is achieved by performing gradient descent on the loss function. In this embodiment, the loss function is as follows:

[0092]

[0093] where L g (u z ) is the loss function, θ is the learning rate, u z , u v are the feature vectors of the traffic network graph data, T is the learning parameter, is the probability distribution. Use the gradient update method to continuously update L g (u z ) until convergence. Based on the loss function provided in this embodiment, adjacent vertices can have similar vector representations, while the representations of separated vertices have a large degree of discrimination.

[0094] Step (IV) and step (V) are completed by the graph convolutional neural network, mainly implemented through the graph convolutional neural network structure model. During the propagation process, the target node will continuously update the node's feature representation, generalize and update the node's feature sequence, and change the label category of each node. Finally, through the label type of each node, the automatic allocation of traffic resources can be completed, and a reasonable allocation node sequence, that is, the target topological structure, can be given, providing scientific support for the optimal decision-making of traffic resources.

[0095] Specifically, taking a certain transportation network G as an example, a traffic resource allocation operation is performed. First, assume that the label type of a certain site in the transportation network G is 2, which indicates that the resources of the node v corresponding to this site are moderate. After a round of resource reallocation, the resource allocation situation of the transportation network G is updated, and thus the node feature matrix is updated. At this time, the graph convolutional neural network model starts a new round of node propagation, node feature aggregation, and node feature update process, and the specific process is as follows: Update the feature matrix of all traffic nodes in G according to the change of resources of each node in the transportation network G; When the feature matrix of the traffic node is updated, the graph convolutional neural network model performs a Fourier transform on the updated transportation network graph and propagates between the node and its neighbor nodes. During this process, the features of the target node are aggregated according to the features of the neighbor nodes; After the aggregation ends, the label type of the target node is updated according to the aggregation result. At this time, the backpropagation process starts, and the label types of the neighbor nodes of the target node are also updated accordingly. When a round of label update ends, the label type of each traffic site represents the current tightness of its resources. If the name of a traffic node with a label of 1 is given, it means that the resources of this traffic node are in short supply, and its resources need to be automatically allocated; If the label type of node v is updated from 2 to 1, it indicates that the resources of node v are in short supply, and thus the resource allocation work for the traffic site corresponding to this node v is automatically performed.

[0096] In an embodiment of the present invention, a traffic resource allocation system based on a belief graph convolutional neural network is provided. The traffic resource allocation system includes a modeling unit, a feature extraction unit, and a data processing unit. Among them, the modeling unit is configured to model the current actual transportation network and traffic sites to generate traffic network graph data; The feature extraction unit is configured to extract features from the traffic network graph data generated by modeling through a graph embedding method to determine a feature extraction vector; The data processing unit is configured to process the feature extraction vector based on a deep belief network and propagate and update it through a graph convolutional neural network, thereby updating the feature representation of the node and changing the label type of each node to generate a target topological structure, and then automatically allocate traffic resources to the corresponding traffic sites according to the target topological structure to achieve an automatic allocation of traffic resources. Specifically, the specific working mode of this traffic resource allocation system is combined with Figure 1 and with reference to the specific description of the above traffic resource allocation method in this text, it will not be elaborated here.

[0097] The idea of the embodiment of this traffic resource allocation system is the same as the working process of the traffic resource allocation method in the above embodiment. The entire content of the above embodiment of the traffic resource allocation method is incorporated into the embodiment of this traffic resource allocation system by way of full reference and will not be elaborated here.

[0098] In an embodiment of the present invention, a storage medium is provided. The storage medium is configured to store a computer program, and the computer program is configured to execute the traffic resource allocation method described above.

[0099] The idea of the storage medium embodiment is the same as the working process of the traffic resource allocation method in the above embodiment. The entire content of the traffic resource allocation method embodiment is incorporated into this storage medium embodiment by way of full reference and will not be repeated here.

[0100] The present invention takes graph data as input and, based on the high-precision modeling of the real environment, uses the confidence graph convolutional neural network that combines the deep belief convolutional neural network and the graph convolutional neural network to achieve the automatic allocation of traffic resources. By providing a modeling method for graph data facing real-world dynamic changes, the dynamic changes of all traffic resources are abstracted into vector representations, and the intelligent automatic allocation of traffic resources is achieved through the calculation and fitting of neural networks. Specifically, the confidence graph convolutional neural network is used to dynamically capture the resource changes in the real traffic network and, based on this, achieve the automatic allocation of traffic resources. The present invention mainly captures the traffic network resources through traffic network abstraction, that is, creates an abstract graph structure with traffic stations as nodes, the flow of traffic resources between traffic stations is abstracted as edges, the resource distribution is assigned as node features, and the resource changes are captured through the changes in node features. The label type is set according to the degree of scarcity of traffic resources. The confidence graph convolutional neural network performs learning and prediction, and can update the node labels according to the neighbor node feature matrix of the target node, thereby updating the resource distribution of traffic stations.

[0101] The above are only the preferred embodiments of the present invention, and do not limit its patent scope accordingly. Any equivalent structural or equivalent process transformation made by using the specification and drawings of the present invention, directly or indirectly applied in other related technical fields, shall be similarly included in the patent protection scope of the present invention.

Claims

1. A traffic resource allocation method based on a confidence graph convolutional neural network, characterized in that The described traffic resource allocation method is used for the automatic allocation of traffic resources. The traffic resource allocation method includes: Modeling the current actual traffic network and traffic stations, the process of which includes converting the actual traffic network to generate an edge set of traffic network graph data, and converting some traffic stations to generate a node set of traffic network graph data; Feature extraction is performed on the traffic network graph data generated by modeling through a graph embedding method to determine a feature extraction vector; the process of feature extraction includes obtaining the original feature representation of the node. The original feature of the node is obtained by repeatedly starting from any node in the graph data and transferring between nodes with equal probability to collect multiple groups of node sequences, and inputting the original feature representation into a Skipgram model for processing to obtain the final feature extraction vector; during backpropagation, part of the weights of the hidden layer are updated, and it is judged whether to update the weights according to the node appearance probability related to the weights. The greater the node appearance probability, the greater the probability of updating the weights related to it, and continue to sample some nodes, and control the probability of sampling the nodes. In part of the sampling, the higher the frequency of the node appearance, the smaller the probability of using it as the starting point for sampling; Based on a deep belief network, the feature extraction vector is processed, and propagated and updated through a graph convolutional neural network, so as to update the feature representation of the node and change the label type of each node to generate a target topological structure, and then automatically allocate traffic resources to the corresponding traffic stations according to the target topological structure to achieve an automatic allocation of traffic resources; the label type of the node represents the current traffic resource tension degree of the traffic station corresponding to the node. When the tension degree is higher, more traffic resources are allocated to the corresponding traffic station; The processing process based on the deep belief network includes unsupervised pre-training based on RBM and weight initialization through the CD-k method, so as to calculate the probability distribution of the hidden layer confidence neurons in the deep belief network being activated, and then update the weights based on the probability distribution; calculate the probability distribution of the hidden layer confidence neurons in the deep belief network being activated according to the following formula: In the formula, is the probability distribution, v (0) is the set of nodes from the traffic network diagram data, θ is the learning rate, W j is the vector matrix to be processed, b j is the bias; The process of propagation and update based on the graph convolutional neural network includes aggregating the node set of the traffic network graph data through pooling propagation; After completing an automatic allocation of traffic resources, update the resource allocation situation of the current actual traffic network, and perform an automatic allocation of traffic resources again based on the updated actual traffic network data.

2. The traffic resource allocation method based on a confidence graph convolutional neural network according to claim 1, characterized in that The traffic resource allocation method further includes: The method for updating the resource allocation situation of the current actual traffic network includes: Update the feature matrix of all traffic nodes according to the change situation of traffic resources at each traffic station in the actual traffic network after the automatic allocation to obtain new traffic network graph data; Use the GCN network model to perform Fourier transform on the new traffic network graph data and propagation between the node and its neighbor nodes, where the features of the target node related to the target topological structure are aggregated according to the features of its neighbor nodes; Update the label type of the target node according to the aggregation result, and at the same time start backpropagation to update the label types of the neighbor nodes of the target node.

3. The traffic resource allocation method based on a confidence map convolutional neural network according to claim 1, wherein The method for obtaining the original feature representation of a node includes: Repeat the following steps to collect multiple sets of node sequences: start from any node in the graph data and transfer between the nodes with equal probability.

4. The traffic resource allocation method based on a confidence graph convolutional neural network according to claim 1, wherein The method for inputting the original feature representation into the Skipgram model to obtain the final feature extraction vector includes: The probability of the weight being updated is determined according to the following formula: Where P(v) vi is the occurrence probability of node v i , Z(v i ) is the weight of node v i , Z(v j ) is the weight of node v j , and v i , v j are from the node set of the traffic network diagram data; The node occurrence probability is determined according to the following formula: Wherein, P(o i ) is the occurrence probability of node o in the network constructed according to the existing weights i , and Z(o i ) is the weight of node o in the network constructed according to the existing weights i .

5. The traffic resource allocation method based on a confidence graph convolutional neural network according to claim 1, wherein The processing process based on the deep belief network further includes: Continue Gibbs sampling according to the calculated probability distribution, and extract corresponding values for the neurons in the visible layer of the deep belief network for sampling reconstruction, so as to update the new weights and biases, and the process is represented by the following formula: W ← W + λ[x(h (0) = 1)|z (0) z (0)T -x(x (1) = 1|v(1))z (1)T where W is the new weight matrix, λ and T are learning parameters, x is the sampling probability, and h (0) is the hidden layer representation of the 0th layer, and the calculation method of the h (0) includes, after the aggregation operation at the 0th layer, adding the results obtained by the nodes on different subgraphs. z (0) is the output at the node level of the 0th layer, x (1) is the sampling probability of the 1st layer, z (1) is the output at the node level of the 1st layer, and v(1) is the node set of the 1st layer.

6. The traffic resource allocation method based on a confidence map convolutional neural network according to claim 1, wherein Aggregate the node set of the traffic network graph data according to the following formula: where pool is the pooling operation, s is the vector of the s-th layer, θ is the learning rate, W is the adjacent vector, is the prediction result of node v i , k represents the k-th layer of the neural network, b is the offset, and v i comes from the node set V of the traffic network graph data.

7. The traffic resource allocation method based on a confidence map convolutional neural network according to claim 1, characterized in that The process of updating based on the graph convolutional neural network further includes: Update the feature representation of the node by performing gradient descent on the loss function, where the loss function is determined by the following formula: where, L g (u z ) is the loss function, θ is the learning rate, u z , u v are the feature vectors of the traffic network graph data, T is the learning parameter, is the probability distribution.

8. A traffic resource allocation system based on a confidence graph convolutional neural network, characterized in that The traffic resource allocation system performs traffic resource allocation based on the method according to any one of claims 1 to 7, and it includes: A modeling unit configured to model the current actual traffic network and traffic stations to generate traffic network graph data; A feature extraction unit configured to perform feature extraction on the traffic network graph data generated by modeling through a graph embedding method to determine a feature extraction vector; A data processing unit configured to process the feature extraction vector based on a deep belief network, and propagate and update it through a graph convolutional neural network, so as to update the feature representation of the node and change the label type of each node to generate a target topological structure, and then automatically allocate traffic resources to the corresponding traffic stations according to the target topological structure to achieve automatic allocation of traffic resources at one time.

9. A storage medium, characterized in that, The storage medium is configured to store a computer program, and the computer program is configured to execute the traffic resource allocation method according to any one of claims 1 to 7.

Images