Flow prediction method based on dynamic graph space-time correlation and adaptive adversarial training

Abstract

The invention relates to a traffic prediction method based on dynamic graph space-time correlation and adaptive adversarial training. The method comprises the following steps: acquiring a dynamic graph space embedding representation by learning dynamic graph data capture by utilizing an adaptive graph convolutional network; capturing the dynamic graph data according to the time attention network to obtain a dynamic graph time embedding representation; obtaining dynamic graph embedding; after dynamic graph embedding is used as initial embedding and the last preset adversarial node embedding is defined as context node embedding to be used as input of a strategy network, an adversarial reward function and a loss function are set, then the strategy network is trained, and an updated dynamic graph is output according to the trained strategy network. And performing spatial correlation capture and temporal correlation capture on the updated dynamic graph to obtain a traffic prediction result corresponding to the current network equipment. By adopting the method, data pollution and node abnormity can be effectively handled, and traffic prediction can be realized by adapting to continuously changing network conditions.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence and data analysis technology, and in particular to a traffic prediction method based on dynamic graph spatiotemporal correlation and adaptive adversarial training. Background Art

[0002] Network traffic prediction is a technology that uses historical network traffic data to estimate the size and trend of future network traffic. It is very important in network planning, operation and management, and can help network managers allocate network resources more effectively and improve network performance and reliability. By using network traffic prediction, by estimating the future network load in advance, measures can be taken to solve the problem before traffic overload occurs, thereby improving network availability. This method is more effective than traditional responsive behavior. Frequent data transmission and information exchange are required between a large number of systems distributed in various places. Although the continuous improvement of network infrastructure construction can generally meet the basic requirements of communication, in some local areas and during some periods, there will be a phenomenon that bandwidth requirements cannot be met due to unreasonable resource allocation. Therefore, artificial intelligence and machine learning technologies are used to help communication systems better predict user network traffic needs, optimize resource deployment and execution strategies in advance, improve the overall performance of the system, and effectively respond to sudden demands, improve network stability and service quality, reduce costs and improve efficiency. Nodes of network equipment, such as servers, routers or switches, bear the important responsibilities of receiving, processing and forwarding data. However, in complex network environments, nodes are often disturbed. For example, abnormalities caused by erroneous inputs and malicious attacks can lead to inaccurate traffic prediction, which in turn leads to incorrect resource allocation strategies and reduced response time. Therefore, it is urgent to design a robust traffic prediction method that can maintain high accuracy in various uncertain situations.

[0003] Traditional traffic robust prediction methods are mainly divided into robust statistical methods, machine learning methods, and deep learning methods. Robust statistical methods such as the robust autoregressive moving average method resist the influence of abnormal data by adjusting its error term, and are suitable for processing linear time series containing outliers. In the field of machine learning, ensemble learning methods, such as random forests and gradient boosting machines, integrate the results of multiple prediction models to reduce the influence of single model bias, thereby enhancing the robustness of prediction. In deep learning technology, especially the variants of denoising autoencoders and recurrent neural networks, such as long short-term memory networks with attention mechanisms, can effectively process nonlinear complex patterns in data and provide reliable prediction results even when data quality is unstable. In general, although traditional robust prediction methods can effectively handle abnormal data to a certain extent and ensure the stability of prediction results, these methods often perform poorly when faced with the challenges of dynamic network environments. With the continuous evolution of network attack patterns and the increase in attack types, traditional prediction methods are difficult to adapt quickly, resulting in a decrease in prediction accuracy. In addition, most adversarial training methods are designed for static environments, and they have limited effect in dynamic traffic prediction tasks. In the face of a constantly changing network environment, fixed static strategies are not effective. Summary of the invention

[0004] Based on this, it is necessary to provide a traffic prediction method based on dynamic graph spatiotemporal correlation and adaptive adversarial training, which can effectively deal with data pollution and node anomalies, and adapt to changing network conditions to achieve traffic prediction, in order to address the above technical problems.

[0005] A traffic prediction method based on dynamic graph spatiotemporal correlation and adaptive adversarial training, the method comprising:

[0006] Modeling network devices as dynamic graphs; the dynamic graphs include node sets, edge sets, and adjacency matrices; the node sets include multiple nodes; the nodes include servers, routers, and switches; obtaining dynamic graph data corresponding to network devices at historical moments;

[0007] The adaptive graph convolutional network is used to capture the dynamic changes of nodes and edges in the dynamic graph data to obtain the dynamic graph spatial embedding representation; the temporal correlation between nodes in the dynamic graph data is captured according to the temporal attention network to obtain the dynamic graph temporal embedding representation;

[0008] The dynamic graph spatial embedding representation and the dynamic graph temporal embedding representation input are nonlinearly activated and linearly combined to obtain the dynamic graph embedding; the dynamic graph embedding is used as the starting embedding and the pre-set last adversarial node embedding is defined as the context node embedding as the input of the policy network;

[0009] In the policy network, the context nodes are updated according to the multi-head attention mechanism to obtain the updated initial node embedding; the probability of the next node is calculated using the updated initial node embedding, and the node with the highest probability is selected as the next sampled adversarial node among all nodes; the adversarial node is applied to the policy network, and the adversarial reward function and loss function of the policy network are set according to the change in prediction performance before and after the adversarial node is applied;

[0010] The strategy network is trained using an adversarial reward function and a loss function, and an updated adversarial node is output according to the trained strategy network; a dynamic graph is updated according to the updated adversarial node, and spatial correlation and temporal correlation are captured on the updated dynamic graph to obtain a traffic prediction result corresponding to the current network device.

[0011] The above-mentioned traffic prediction method based on spatiotemporal correlation and adaptive adversarial training of dynamic graphs, this application models the network device as a dynamic graph, and uses an adaptive graph convolutional network to capture the dynamic changes of nodes and edges in the dynamic graph data to obtain a dynamic graph spatial embedding representation; according to the temporal attention network, the temporal correlation between nodes in the dynamic graph data is captured to obtain a dynamic graph temporal embedding representation, and the adaptive graph convolutional network is used to effectively mine the complex spatial interaction patterns between nodes, while the temporal attention network keenly captures the dynamic evolution characteristics of time series traffic. The two collaborate to gain insight into and predict the dynamic changes of traffic patterns, effectively improving the accuracy of traffic data prediction. Furthermore, the present application designs a strategy generation mechanism, which performs nonlinear activation and linear combination on the dynamic graph spatial embedding representation and dynamic graph temporal embedding representation input to obtain dynamic graph embedding; defines the dynamic graph embedding as the starting embedding and the pre-set last adversarial node embedding as the context node embedding as the input of the policy network; updates the context node in the policy network according to the multi-head attention mechanism, calculates the next node probability using the updated initial node embedding, and selects the node with the highest probability among all nodes as the next sampled adversarial node to automatically generate an adversarial node sequence, which represents a specific set of nodes used to generate adversarial samples, simulating the network's response when encountering hypothetical attacks or failures. By applying the projected gradient descent method, adversarial samples play a "stress test" function in the training cycle, constantly tempering the method's anti-interference strength. Finally, by setting the policy network loss function to systematically evaluate and fine-tune the performance of the nodes selected by the policy network, the overall adversarial training framework is further iterated and optimized to ensure that the traffic prediction method maintains a high degree of stability and predictability in the rapidly changing network ecology. This application uses a dynamic node selection strategy for adversarial training to solve the problem of insufficient robustness of current traffic prediction methods in the face of data pollution and node anomalies, while improving the accuracy and stability of predictions through adaptive graph convolutional networks and temporal attention mechanisms. On the other hand, this application uses real-time traffic data input to output future traffic trends and potential anomaly warnings to guide network operation and maintenance and policy tuning, covering flexible allocation of bandwidth configuration, intelligent optimization of routing strategies, and pre-identification and mitigation of network risks to ensure uninterrupted and secure network services. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A flow chart of a traffic prediction method based on dynamic graph spatiotemporal correlation and adaptive adversarial training in one embodiment;

[0013] Figure 2 A schematic diagram of a network traffic monitoring, analysis and prediction process in one embodiment;

[0014] Figure 3 A schematic diagram of a strategy neural network structure diagram in one embodiment;

[0015] Figure 4 It is a schematic diagram of the flow structure of traffic prediction based on dynamic graph spatiotemporal correlation and adaptive adversarial training in another embodiment;

[0016] Figure 5 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0018] In one embodiment, Figure 1 As shown, a traffic prediction method based on dynamic graph spatiotemporal correlation and adaptive adversarial training is provided, including the following steps:

[0019] Step 102, modeling the network device as a dynamic graph; the dynamic graph includes a node set, an edge set and an adjacency matrix; the node set includes multiple nodes; the nodes include servers, routers and switches; and obtaining dynamic graph data corresponding to the network devices at historical moments.

[0020] In this application, servers, routers, and switches are regarded as nodes in the diagram. They are the cornerstones of the network architecture and play a decisive role in the control of data flow. Traffic data, reflected in the amount and frequency of data packet transmission on these nodes, is a direct indicator for evaluating the active status of the network. It is collected in real time through SNMP and NetFlow, and comprehensively covers information such as bandwidth usage, upstream and downstream rates, and packet size. The robust traffic prediction method proposed in this application uses real-time traffic data input to output future traffic trends and potential anomaly warnings to guide network operation and maintenance and policy tuning, covering flexible allocation of bandwidth configuration, intelligent optimization of routing strategies, and pre-identification and mitigation of network risks to ensure the uninterrupted and secure network services. Figure 2 As shown, a systematic process of traffic analysis and prediction is demonstrated.

[0021] Step 104, using an adaptive graph convolutional network to capture the dynamic changes of nodes and edges in the dynamic graph data by learning to obtain a dynamic graph spatial embedding representation; capturing the temporal correlation between nodes in the dynamic graph data according to a temporal attention network to obtain a dynamic graph temporal embedding representation.

[0022] The dynamic graph can reflect the changes in network status in real time, including the update of node connectivity and node characteristics, so that the method can adapt to the complex and changing network environment, thereby improving the accuracy of prediction and the response speed of the method. The traffic data is represented as a dynamic graph G = (V, E, A), where V is a node set, representing each node in the network, and |V| = N represents the number of nodes in the node set V; E is a set of edges, representing the connection relationship between nodes, is the adjacency matrix. The dynamic graph can dynamically reflect the information changes of nodes and edges in the network, especially by continuously updating in time series, thus truly simulating the dynamic behavior of the actual network. Define the function F θ , to predict the next τ steps of data based on the past T steps of historical data:

[0023] {X t+1 ,X t+2 ,…,X t+τ =F θ (X t ,X t-1 ,…,X t-T+1 ; G)

[0024] Among them, X t represents the flow data at time t, {X t+1 ,X t+2 ,...,X t+τ} represents the predicted traffic data from time t+1 to time t+τ. Prediction function F θ Using the traffic data of the past T steps {X t ,X t-1 ,...,X t-T+1} and the dynamic graph G to predict the traffic data for the next τ steps, and θ represents the learnable parameters in the method.

[0025] Adaptive graph convolutional neural network is used to optimize the spatial dependency between nodes in dynamic graphs and reduce the number of parameters. Adaptive graph convolutional neural network captures the dynamic characteristics of network structure by learning the dynamic changes of nodes and edges in dynamic graphs, thus more accurately reflecting the actual state of the network. Two parameter matrices are learned: node embedding matrix and the weight matrix Where d is the embedding dimension. Through matrix decomposition, the weight matrix is ​​represented as M AS ·W AS , the formula is as follows:

[0026]

[0027] In order to fully reflect the spatial correlation of the data, the dependencies between the nodes of the dynamic graph are adaptively inferred. Randomly initialize the node embedding matrix Where R AEach row represents the embedding of a node, and d represents the embedding dimension of the node. The spatial dependency between nodes in the dynamic graph is inferred by the following formula:

[0028]

[0029] Among them, ReLU is the activation function, represents the Hadamard product, M is used to adjust the weight of each element, B is the bias matrix, and the adaptive graph matrix R A It is automatically updated during training to learn node features in dynamic graphs.

[0030] By optimizing the spatial dependencies between nodes in dynamic graphs through adaptive graph convolutional networks, we can gain a deep understanding of how nodes in the network influence each other and how these interactions change, thereby more accurately predicting traffic data.

[0031] In dynamic graphs, traffic data not only has spatial dependence between nodes, but also has strong temporal correlation, especially traffic at similar times often has similar patterns. In order to better capture these spatiotemporal characteristics, this application uses a temporal attention network to capture the temporal correlation of dynamic graphs. The temporal attention network includes an encoding module, a graph attention module, a multi-layer capture module, a feature reconstruction module, an important feature matrix module, and a feature learning module; according to the encoding module, the input dynamic graph data is encoded to obtain a node feature matrix and a timestamp node feature matrix; in the graph attention module, a fused attention matrix is ​​calculated based on the input node feature matrix and the timestamp node feature matrix; in the multi-layer capture module, the node features of the next moment are captured according to the fused attention matrix at the current moment; in the feature reconstruction module, the captured node feature matrix is ​​reconstructed, and the node features at the current moment are supplemented by calculating the difference between the features of the static attention output of the graph and the features of the dynamic attention output of the graph, and the reconstructed features are obtained; in the important feature matrix module, the reconstructed features are processed according to the self-attention mechanism to obtain an important feature matrix; in the feature learning matrix, the important feature matrix is ​​feature-learned to obtain a dynamic graph temporal embedding representation.

[0032] The temporal attention mechanism can better capture and predict traffic changes in dynamic networks by analyzing the temporal correlation between nodes in dynamic graphs. This process can understand and utilize patterns in time series, thereby improving the ability to capture dynamic network behaviors.

[0033] Step 106, nonlinearly activate and linearly combine the dynamic graph spatial embedding representation and the dynamic graph temporal embedding representation input to obtain the dynamic graph embedding; define the dynamic graph embedding as the starting embedding and the pre-set last adversarial node embedding as the context node embedding as the input of the policy network.

[0034] Step 108, in the policy network, the context nodes are updated according to the multi-head attention mechanism to obtain the updated initial node embedding; the probability of the next node is calculated using the updated initial node embedding, and the node with the highest probability is selected as the next sampled adversarial node among all nodes; the adversarial node is applied to the policy network, and the adversarial reward function and loss function of the policy network are set according to the change in the predicted performance before and after the adversarial node is applied.

[0035] In this application, the policy network is designed to dynamically adapt to the time evolution of the graph structure and specifically select key nodes that may significantly affect the prediction performance of the method when attacked or contaminated. The core of the policy network is a multi-head attention mechanism that is optimized to process the real-time features of nodes in the dynamic graph, so as to more accurately evaluate the criticality of each node. Its structure is as follows Figure 3 As shown. The dynamic graph spatial embedding representation and the dynamic graph temporal embedding representation are activated nonlinearly and then linearly combined to obtain the final dynamic graph embedding, that is:

[0036] U=c 1 ·Φ 1 (O)+(1-c 1 )·Φ 2 (h)

[0037] Among them, Φ 1 , Φ 2 is a learnable parameterized activation function, c 1 Balance the weights of spatial and temporal embedding.

[0038] The input of the policy network is the context node embedding, including the starting embedding and the last adversarial node embedding. The context node embedding is defined as follows:

[0039]

[0040] where v is a learned embedding in the first iteration step, is the embedding of the last selected adversarial node in the k-1th iteration step.

[0041] To update the context node embedding, the multi-head attention mechanism is calculated as follows:

[0042]

[0043] Among them, ∥ represents the serial operation, J is the number of heads of multi-head attention, MHA j It is the multi-head attention mechanism of the j-th head, and the calculation formula is as follows:

[0044]

[0045] Query vector q in attention mechanism (c) , key vector k j Sum value vector v j is defined as follows:

[0046] k j =W j K H′,v j =W j V H′(j=1,…,M)

[0047] in, W j K , W j V is the linear transformation matrix, and H′ is the output of the previous layer.

[0048] To calculate the probability of the next node, we first calculate the query vector q and key vector k based on the initial node embedding i :

[0049] q=W Q U (c) ′,k i =W K H i '

[0050] First, we compute the logit with a single-headed attention, querying all nodes with the updated context node:

[0051]

[0052] Where C is a constant, d k is the dimension of the key vector, r j is the attention score of node j. The selected node is represented by r j = -∞ or the smallest negative number that the computer can express to mask. Then use the softmax function to calculate the final probability of the node:

[0053]

[0054] Among them, p i is the probability of node i, ω k is the current node, s is the current state, ω 1:k-1 is the node sequence selected previously, r i is the attention score of node i. The node with the highest probability is selected as the next sampled adversarial node among all nodes, ensuring that the policy network not only considers the immediate characteristics of the node, but also evaluates its potential threat to the stability of the network during the dynamic evolution of the entire network.

[0055] Through the dynamic graph-based policy network, when the prediction performance may be threatened, the system's defense capability is enhanced by targeted selection of key nodes, thereby improving the robustness of the entire network.

[0056] In order to evaluate the accuracy of the policy network when generating adversarial nodes, this application uses internal loss (the change in prediction performance before and after applying adversarial nodes) as a reward. Since dynamic graphs can reflect the information changes of nodes and edges in the network in real time, this application uses this feature to make a more accurate evaluation. The larger the loss value, the better the generated adversarial node effect. The calculation formula is as follows:

[0057]

[0058] in is the mean square error loss function, X′ t is the historical data after adversarial perturbation, and G is a dynamic graph that captures the changes of nodes and edges in the network through dynamic updates.

[0059] As training progresses, the method gradually becomes more robust and the internal loss is expected to decrease. This may lead to incorrect feedback and optimization jitter. In order to improve the adversarial training effect, knowledge distillation is used to transfer the knowledge of the teacher network to the student network. By exploiting the spatiotemporal characteristics captured by the dynamic graph, the previous round of training network is used as the teacher network and the current network to be trained is used as the student network. The knowledge distillation reward is defined as follows:

[0060]

[0061] Among them, L KD is the knowledge distillation loss, is the mean square error loss function, X′ t is the historical data after adversarial disturbance, is the teacher network trained last time.

[0062] In order to enhance the learning of the policy network, the Markov distance between the current network and the teacher network is calculated to increase the training reward. The formula is as follows:

[0063]

[0064] Among them, T K (d) shows the method of calculating distance, which takes into account the dynamic changes of nodes and edges in the dynamic graph.

[0065] Using the Markov distance as part of the reward ensures that the distance between the target policy network and the teacher network is as small as possible. The final adversarial reward is:

[0066]

[0067] Where c is a constant, α 1 and α 2 are parameters that control the amount of knowledge transferred from the teacher network and the weight of the Markov distance reward.

[0068] By utilizing the dynamic characteristics of dynamic graphs, the optimal adversarial nodes can be effectively evaluated and selected, the robustness of the network in the face of data pollution and node anomalies can be enhanced, and the prediction performance and stability of dynamic graphs can be ensured.

[0069] Finally, the policy network is trained to optimize the dynamic graph prediction performance. The policy network is trained by alternately training the policy network and the traffic prediction method. The policy network is trained according to the input X t-τ:t Generate an adversarial node sequence, denoted as Ω. Then calculate the reward of the corresponding operation and use the reward to update the policy network. Next, use the selected adversarial node sequence to calculate the adversarial sample X′ through the projected gradient descent method t-τ:t′ :

[0070]

[0071] The constraints are:

[0072]

[0073] Among them, X t+1:t+τ ={X t+1 ,X t+2 ,…,X t+τ} is the next τ step data, ∈ is a predefined perturbation budget, which limits the maximum change between the adversarial sample and the original data to ensure that the generated adversarial sample is within a reasonable range. The adversarial sample is generated by X at the i-th iteration. t -τ:t ′(i) Indicates that γ is the step size, I t ∈{0,1} n is the final node selection metric obtained from the policy network, is the mean square error loss function, and sign is the sign function, indicating the gradient direction.

[0074] The traffic prediction method then uses these adversarial examples to optimize the prediction loss:

[0075]

[0076] To train the policy network, the loss function is defined as follows:

[0077]

[0078] Where c is a constant, s represents the problem instance, which consists of traffic data from time point t-T+1 to time point t. The policy network is optimized by gradient descent using the Adam optimizer. The detailed operation steps of the traffic robust prediction method are as follows. Figure 4 , adversarial node sequences are generated through policy networks, adversarial samples are generated using the projected gradient descent method, and the training process is optimized through internal loss and knowledge distillation rewards, so that dynamic graphs can adapt to complex and dynamic application environments and improve the defense capabilities of dynamic graphs against dynamic attacks.

[0079] Step 110, train the policy network using the adversarial reward function and the loss function, output an updated adversarial node based on the trained policy network; update the dynamic graph based on the updated adversarial node, capture the spatial correlation and the temporal correlation of the updated dynamic graph, and obtain the traffic prediction result corresponding to the current network device.

[0080] The strategy network is trained using an adversarial reward function and a loss function, and an updated adversarial node is output according to the trained strategy network; the dynamic graph is updated according to the updated adversarial node, and the adversarial node in the initial dynamic graph is replaced with the updated adversarial node to obtain an updated dynamic graph, and then the spatial correlation and temporal correlation of the updated dynamic graph are captured, and the captured dynamic graph spatial embedding representation and temporal embedding representation are passed through a two-dimensional convolution layer to obtain the traffic prediction result corresponding to the current network device. In the above-mentioned traffic prediction method based on dynamic graph spatiotemporal correlation and adaptive adversarial training, the present application obtains the dynamic graph spatial embedding representation by modeling the network device as a dynamic graph, and uses an adaptive graph convolution network to capture the dynamic changes of nodes and edges in the dynamic graph data by learning; the temporal correlation between nodes in the dynamic graph data is captured according to the temporal attention network to obtain the temporal embedding representation of the dynamic graph, and the complex spatial interaction patterns between nodes are effectively mined through the adaptive graph convolution network, while the temporal attention network keenly captures the dynamic evolution characteristics of the time series traffic. The two collaborate to understand and predict the dynamic changes of the traffic pattern, effectively improving the accuracy of traffic data prediction. Furthermore, the present application designs a strategy generation mechanism, which performs nonlinear activation and linear combination on the dynamic graph spatial embedding representation and dynamic graph temporal embedding representation input to obtain dynamic graph embedding; defines the dynamic graph embedding as the starting embedding and the pre-set last adversarial node embedding as the context node embedding as the input of the policy network; updates the context node in the policy network according to the multi-head attention mechanism, calculates the next node probability using the updated initial node embedding, and selects the node with the highest probability among all nodes as the next sampled adversarial node to automatically generate an adversarial node sequence, which represents a specific set of nodes used to generate adversarial samples, simulating the network's response when encountering hypothetical attacks or failures. By applying the projected gradient descent method, adversarial samples play a "stress test" function in the training cycle, constantly tempering the method's anti-interference strength. Finally, by setting the policy network loss function to systematically evaluate and fine-tune the performance of the nodes selected by the policy network, the overall adversarial training framework is further iterated and optimized to ensure that the traffic prediction method maintains a high degree of stability and predictability in the rapidly changing network ecology. This application uses a dynamic node selection strategy for adversarial training to solve the problem of insufficient robustness of current traffic prediction methods in the face of data pollution and node anomalies, while improving the accuracy and stability of predictions through adaptive graph convolutional networks and temporal attention mechanisms. On the other hand, this application uses real-time traffic data input to output future traffic trends and potential anomaly warnings to guide network operation and maintenance and policy tuning, covering flexible allocation of bandwidth configuration, intelligent optimization of routing strategies, and pre-identification and mitigation of network risks to ensure uninterrupted and secure network services.

[0081] In one embodiment, an adaptive graph convolutional network is used to capture the spatial correlation of a dynamic graph by learning the dynamic changes of nodes and edges in the dynamic graph data, including:

[0082] The adaptive graph convolutional network is used to capture the dynamic changes of nodes and edges in the dynamic graph data and obtain the spatial correlation of the dynamic graph:

[0083]

[0084] Where N is the total number of nodes in the dynamic graph, σ is the Sigmoid activation function, i is the current node, j is its neighbor node, k represents all other nodes that may affect node i, and I N is the unit matrix, I represents the input dynamic graph data, O represents the output dynamic graph spatial correlation, is the adjacency matrix, D is the degree matrix, is the node embedding matrix, is the weight pool, is the bias pool, where d is the embedding dimension, C is the feature dimension, F is the output feature dimension, and ReLU is the activation function. represents the Hadamard product, M is used to adjust the weight of each element, B is the bias matrix, and R A is an adaptive graph matrix, which is automatically updated during the training process. exp is an exponential function used to calculate the exponential power of e.

[0085] In one embodiment, the temporal attention network includes an encoding module, a graph attention module, a multi-layer capture module, a feature reconstruction module, an important feature matrix module and a feature learning module; the temporal correlation between nodes in the dynamic graph data is captured according to the temporal attention network to obtain a dynamic graph temporal embedding representation, including:

[0086] Encode the input dynamic graph data according to the encoding module to obtain a node feature matrix and a timestamp node feature matrix;

[0087] In the graph attention module, the fused attention matrix is ​​calculated based on the input node feature matrix and the timestamp node feature matrix;

[0088] In the multi-layer capture module, the fused attention matrix is ​​updated to obtain an updated node matrix;

[0089] In the feature reconstruction module, the updated node matrix is ​​reconstructed, and the node features at the current moment are supplemented by calculating the difference between the features of the graph static attention output and the features of the graph dynamic attention output to obtain the reconstructed features;

[0090] In the important feature matrix module, the reconstructed features are processed according to the self-attention mechanism to obtain the important feature matrix;

[0091] Feature learning is performed on the important feature matrix in the feature learning matrix to obtain the temporal embedding representation of the dynamic graph.

[0092] In a specific embodiment, in order to better capture time correlation, the present application adopts a combination of graph static attention and graph dynamic attention to highlight the importance of changes in node attributes at different time points. Graph static attention is used to capture the fixed attributes of nodes, while graph dynamic attention is used to capture changes in node attributes over time. This design can simultaneously consider the attributes of nodes at a specific time point and the attributes that change over time, thereby improving the performance of the method at different times and enhancing the ability to process dynamic graph data.

[0093] In order to further capture time dependencies, this application uses a multi-layer structure to capture dependencies at different time points:

[0094]

[0095] But in most scenarios, the graph static attention matrix A static is 0, so the above formula can also be written as

[0096]

[0097] Among them, H (l) is the node feature matrix of the lth layer. This design can capture the spatiotemporal dependency of node features at different times, thereby improving the performance of the prediction method in dynamic graph data processing.

[0098] In order to further enhance the feature representation capability of the method, this application introduces a variable Transformer for feature reconstruction. The variable Transformer is a complex network composed of different modules that can sample features of traffic data at important time points and process these features through classic multi-head attention layers. First, key time points are selected and feature sampling of traffic data at these time points is performed. Then, these features are processed using a multi-head attention mechanism, the features output by each attention head are spliced ​​together, and adjusted through linear transformations and offsets. The core is to compare the difference between the features output by the static attention of the graph and the features output by the dynamic attention of the graph:

[0099] Q=[Q 1 ,Q 2 ,…,Q h ]

[0100] K=[K 1 ,K 2 ,…,Kh ]

[0101] V=[V 1 ,V 2 ,…,V h ]

[0102] head i =Attention(Q i ,K i ,V i )

[0103]

[0104] ΔF=F static -F dynamic

[0105] Among them, F static and F dynamic are the output features of static and dynamic attention, Q, K, V are the matrices of query, key and value, respectively, and W O is the output linear transformation matrix. This difference ΔF is assigned to the currently selected time map data to adjust and compensate for the differences between the features, thereby accurately reflecting the time dependency. This method not only improves the prediction ability, but also enhances the ability to capture and process the features of key time points.

[0106] Finally, the relationship matrix, that is, the important feature matrix, is determined by summarizing the key features of the entire sequence into a set to capture the relationship between the features, as follows:

[0107] Q=FW Q ,K=FW K ,V=FW V

[0108]

[0109] F att =LayerNorm(F+A)

[0110] Among them, F represents the input feature, W Q , W K and W V is the weight matrix of query, key and value, A is the attention matrix, and LayerNorm represents the layer normalization operation.

[0111] The information in these relationship matrices is applied to each query point in a weighted manner. This design effectively fuses information from a category perspective, and ultimately applies the weighted results to specific tasks, improving the prediction accuracy and robustness of the method, especially when dealing with complex time series data, providing higher method performance and more accurate results.

[0112] In order to enhance the accuracy of the prediction method, the temporal attention mechanism is used to capture the long-term dynamic temporal relationship of traffic data in the dynamic graph. The formula is as follows:

[0113] E=V e ·σ(((x′ i,t ) T W 1 )W 2 (x′ i,t W 3 )+b e )

[0114] Where V e ,b e ,W 1 ,W 2 ,W 3 are all learnable parameters, x′ i,t Represents the input data x at time t i The result after normalization.

[0115] The temporal attention mechanism can better capture and predict traffic changes in dynamic networks by analyzing the temporal correlation between nodes in dynamic graphs. This process can understand and utilize patterns in time series, thereby improving the ability to capture dynamic network behaviors.

[0116] In one embodiment, in the graph attention module, a fused attention matrix is ​​calculated based on the input node feature matrix and the timestamp node feature matrix, including

[0117] In the graph attention module, the graph static attention matrix is ​​calculated based on the input node feature matrix:

[0118] A static =softmax(HW static H T )

[0119] Where H is the node feature matrix, the superscript T represents the transposition operation, and W static Represents the weight matrix of static features, and softmax represents the normalized exponential function, which is used to convert the vector into a probability distribution;

[0120] The graph dynamic attention matrix is ​​calculated based on the input timestamp node feature matrix:

[0121]

[0122] Among them, W dynamic represents the dynamic weight matrix;

[0123] The graph static attention matrix and the graph dynamic attention matrix are fused by weighted addition to obtain the fused attention matrix:

[0124] A combined =αA static +(1-α)A dynamic

[0125] Among them, α is a weight parameter used to control the contribution of static and dynamic attention mechanisms to the fused attention matrix.

[0126] In one embodiment, the fused attention matrix is ​​updated in the multi-layer capture module to obtain an updated node matrix, including:

[0127] The fused attention matrix is ​​updated in the multi-layer capture module, and the updated node matrix is ​​obtained as

[0128]

[0129] Among them, A combined represents the fusion attention matrix, ReLu is the activation function, is the normalized degree matrix, H (l) is the node feature matrix of the lth layer, is the feature weight matrix of the lth layer.

[0130] In one embodiment, feature learning is performed on an important feature matrix in a feature learning matrix to obtain a dynamic graph temporal embedding representation, including:

[0131] In the feature learning matrix, feature learning is performed on the important feature matrix, and the dynamic graph time embedding is obtained as

[0132]

[0133] Among them, F att represents the important feature matrix, R A represents the adaptive graph matrix, E represents the long-term dynamic time relationship of traffic data in the dynamic graph, tanh is the activation function, [·] represents the connection operation, ⊙ represents the element-by-element product operation, z and r are the reset gate and update gate respectively, t represents the current time, σ represents the Sigmoid activation function, represents the candidate hidden state at the current moment, h t-1 Represents the dynamic graph time embedding representation of the previous moment of the current moment, softmax represents the normalized exponential function, which is used to convert the vector into a probability distribution. is an intermediate variable, representing the dependencies between the inferred nodes, W z , W z , bz , b r as well as Represents different learnable parameters, ReLU is the activation function, and the superscript T represents the transposition operation.

[0134] In one embodiment, updating the context node in the policy network according to the multi-head attention mechanism to obtain an updated node embedding includes:

[0135] In the policy network, the context nodes are updated according to the multi-head attention mechanism, and the updated node embedding is obtained as

[0136]

[0137] Among them, ∥ represents the serial operation, J is the number of heads of multi-head attention, MHA j It is the multi-head attention mechanism of the j-th head, and the calculation formula is as follows:

[0138]

[0139] The query vector in the attention mechanism is ≤ (c) , key vector k j Sum value vector v j is defined as follows:

[0140] k i =W j K H′,v j =W j V H′(j=1,…,M)

[0141] Among them, U (c) represents the context node embedding, W j K , W j V is the linear transformation matrix, and H′ is the output of the previous layer.

[0142] In one embodiment, calculating the next node probability using the updated initial node embedding includes:

[0143] Calculate the query vector q and the key vector k of node i based on the initial node embedding i , the formula is as follows:

[0144] q=W Q U (c) ′,k i =W K H i '

[0145] Among them, W Q is the weight matrix for generating query vectors, and W K is the weight matrix for generating key vectors;

[0146] Use single-head attention to calculate the logit to query the updated context node and all nodes:

[0147]

[0148] where C is a constant, d k is the dimension of the key vector, r j is the attention score of node j, tanh is the activation function, the superscript T represents the transpose operation, and k j is the key vector;

[0149] At each step k, select a node ω k , and not only consider the current graph node embedding, but also consider the outputs ω k′ of all previously selected nodes. For all k′ < k, the selected nodes are masked with r j = -∞ or the smallest negative number that the computer can represent, and then use softmax for

[0150]

[0151] where p i is the probability that node i is selected, ω k is the current node, s is the current state, ω 1:k-1 is the sequence of previously selected nodes, r i is the attention score of node i, and p θ represents the probability distribution under the given prediction model parameters θ.

[0152] In one of the embodiments, set the adversarial reward function of the policy network to

[0153]

[0154] where Ω is the set of selected nodes, r(Ω) represents the internal loss, is the mean squared error loss function, X′ t is the historical data after adversarial perturbation at time t, G is the dynamic graph data, r KD is the knowledge distillation reward, and L KD is the knowledge distillation loss, is the teacher network of the previous training, T k (d) represents the method of calculating the distance, represents the Markov distance between the current network and the teacher network, and c is a constant, α 1 and α2 are parameters that control the amount of knowledge transferred from the teacher network and the weight of the Markov distance reward, and the max function is used to find the maximum value in a set of values.

[0155] In one embodiment, the loss function of the policy network is set to

[0156]

[0157] Where c is a constant, s represents the problem instance, which consists of the flow data from time point t-T+1 to time point t, and p φ (Ω|s) represents the probability of selecting the node set Ω under the problem instance s, φ is the parameter of the policy network, log is the logarithmic probability, Represents the gradient.

[0158] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0159] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a traffic prediction method based on dynamic graph spatiotemporal correlation and adaptive adversarial training is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0160] Those skilled in the art will understand that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0161] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0162] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0163] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A traffic prediction method based on dynamic graph spatiotemporal correlation and adaptive adversarial training, characterized in that: The method comprises: The network device is modeled as a dynamic graph; the dynamic graph includes a node set, an edge set and an adjacency matrix; the node set includes a plurality of nodes; the nodes include servers, routers and switches; and the dynamic graph data corresponding to the network device at a historical moment is obtained; The adaptive graph convolution network is used to capture the dynamic changes of nodes and edges in the dynamic graph data to obtain the dynamic graph spatial embedding representation; the temporal correlation between nodes in the dynamic graph data is captured according to the temporal attention network to obtain the dynamic graph temporal embedding representation; The dynamic graph spatial embedding representation and the dynamic graph temporal embedding representation input are nonlinearly activated and linearly combined to obtain a dynamic graph embedding; the dynamic graph embedding is used as a starting embedding and the last pre-set adversarial node embedding is defined as a context node embedding as an input of a policy network; In the policy network, the context node is updated according to the multi-head attention mechanism to obtain an updated initial node embedding; the probability of the next node is calculated using the updated initial node embedding, and the node with the highest probability is selected as the next sampled adversarial node among all nodes; the adversarial node is applied to the policy network, and the adversarial reward function and loss function of the policy network are set according to the change in prediction performance before and after the adversarial node is applied; The strategy network is trained using the adversarial reward function and the loss function, and an updated adversarial node is output according to the trained strategy network; the dynamic graph is updated according to the updated adversarial node, and the spatial correlation and temporal correlation are captured on the updated dynamic graph to obtain a traffic prediction result corresponding to the current network device.

2. The method according to claim 1, characterized in that Adaptive graph convolutional networks are used to capture the spatial correlation of dynamic graphs by learning the dynamic changes of nodes and edges in dynamic graph data, including: The adaptive graph convolutional network is used to capture the dynamic changes of nodes and edges in the dynamic graph data and obtain the spatial correlation of the dynamic graph: Where N is the total number of nodes in the dynamic graph, σ is the Sigmoid activation function, i is the current node, j is its neighbor node, k represents all other nodes that may affect node i, and I N is the unit matrix, I represents the input dynamic graph data, O represents the output dynamic graph spatial correlation, is the adjacency matrix, D is the degree matrix, is the node embedding matrix, is the weight pool, is the bias pool, where d is the embedding dimension, C is the feature dimension, F is the output feature dimension, and ReLU is the activation function. represents the Hadamard product, M is used to adjust the weight of each element, B is the bias matrix, and R A is an adaptive graph matrix, which is automatically updated during the training process. exp is an exponential function used to calculate the exponential power of e.

3. The method according to claim 1, characterized in that The temporal attention network includes an encoding module, a graph attention module, a multi-layer capture module, a feature reconstruction module, an important feature matrix module and a feature learning module; The temporal correlation between nodes in the dynamic graph data is captured according to the temporal attention network to obtain a temporal embedding representation of the dynamic graph, including: Encoding the input dynamic graph data according to the encoding module to obtain a node feature matrix and a timestamp node feature matrix; In the graph attention module, a fused attention matrix is ​​calculated based on the input node feature matrix and the timestamp node feature matrix; In the multi-layer capture module, the fused attention matrix is ​​updated to obtain an updated node matrix; Reconstructing the updated node matrix in the feature reconstruction module, supplementing the node features at the current moment by calculating the difference between the features of the graph static attention output and the features of the graph dynamic attention output, and obtaining the reconstructed features; In the important feature matrix module, the reconstructed features are processed according to the self-attention mechanism to obtain an important feature matrix; Feature learning is performed on the important feature matrix in the feature learning matrix to obtain a dynamic graph temporal embedding representation.

4. The method according to claim 3, characterized in that In the graph attention module, a fusion attention matrix is ​​calculated based on the input node feature matrix and the timestamp node feature matrix, including a graph static attention matrix A calculated based on the input node feature matrix in the graph attention module. static =softmax(HW static H T ) Where H is the node feature matrix, the superscript T represents the transposition operation, and W static Represents the weight matrix of static features, and softmax represents the normalized exponential function, which is used to convert the vector into a probability distribution; The graph dynamic attention matrix is ​​calculated based on the input timestamp node feature matrix: Among them, W dynamic represents the dynamic weight matrix; The static attention matrix and the dynamic attention matrix of the graph are fused by weighted addition to obtain the fused attention matrix A. combined =αA static +(1-α)A dynamic Among them, α is a weight parameter used to control the contribution of static and dynamic attention mechanisms to the fused attention matrix.

5. The method according to claim 3, characterized in that: The fused attention matrix is ​​updated in the multi-layer capture module to obtain an updated node matrix, including: In the multi-layer capture module, the fused attention matrix is ​​updated to obtain the updated node matrix: Among them, A combined represents the fused attention matrix, ReLU is the activation function, is the normalized degree matrix, H (l) is the node feature matrix of the lth layer, is the feature weight matrix of the lth layer.

6. The method according to claim 1, characterized in that Performing feature learning on the important feature matrix in the feature learning matrix to obtain a dynamic graph time embedding representation includes: The feature learning matrix is ​​used to perform feature learning on the important feature matrix, and the dynamic graph time embedding representation is obtained as follows: Among them, F att represents the important feature matrix, R A represents the adaptive graph matrix, E represents the long-term dynamic time relationship of traffic data in the dynamic graph, tanh is the activation function, [·] represents the connection operation, ⊙ represents the element-by-element product operation, z and r are the reset gate and update gate respectively, t represents the current time, σ represents the Sigmoid activation function, represents the candidate hidden state at the current moment, h t-1 Represents the dynamic graph time embedding representation of the previous moment of the current moment, softmax represents the normalized exponential function, which is used to convert the vector into a probability distribution. is an intermediate variable, representing the dependencies between the inferred nodes, W z , W z , b z 、b r as well as Represents different learnable parameters, ReLU is the activation function, and the superscript T represents the transposition operation.

7. The method according to claim 1, characterized in that The context node is updated in the policy network according to the multi-head attention mechanism to obtain an updated node embedding, including: In the strategy network, the context node is updated according to the multi-head attention mechanism, and the updated node embedding is obtained as Among them, ∥ represents the serial operation, J is the number of heads of multi-head attention, MHA j It is the multi-head attention mechanism of the j-th head, and the calculation formula is as follows: Query vector q in attention mechanism (c) , key vector k j Sum value vector v j is defined as follows: Among them, U (c) represents the context node embedding, is the linear transformation matrix, and H′ is the output of the previous layer.

8. The method according to claim 1, characterized in that Calculating the next node probability using the updated initial node embedding includes: Calculate the query vector q and the key vector k of node i based on the initial node embedding i , the formula is as follows: q=W Q U (c) ′,k i =W K H i ′ Among them, W Q is the weight matrix used to generate the query vector, W K is the weight matrix used to generate the key vector; Use single-head attention to calculate logit to query the updated context node and all nodes: Where C is a constant, d k is the dimension of the key vector, r j is the attention score of node j, tanh is the activation function, the superscript T represents the transposition operation, k j is the key vector; At each step k, a node ω is selected k , not only the current graph node embedding needs to be considered, but also the outputs ω of all previously selected nodes k′ , for all k′ < k, the selected nodes are masked with r j = -∞ or the smallest negative number that can be expressed by the computer, and then the softmax function is used to calculate the probability of the next node as Among them, p i is the probability of node i being selected, ω k is the current node, s is the current state, ω 1:k-1 is the node sequence selected previously, r i is the attention score of node i, p θ represents the probability distribution under given prediction model parameters θ.

9. The method according to claim 1, characterized in that: Set the adversarial reward function of the policy network to Where Ω is the selected node set, r(Ω) represents the internal loss, is the mean square error loss function, X′ t is the historical data after adversarial disturbance at time t, G is the dynamic graph data, r KD is the knowledge distillation reward, L KD is the knowledge distillation loss, is the teacher network trained last time, T K (d) represents the distance calculation method, represents the Markov distance between the current network and the teacher network, c is a constant, α1 and α2 are parameters that control the amount of knowledge transferred from the teacher network and the Markov distance reward weight, and the max function is used to find the maximum value in a set of values.

10. The method according to claim 9, characterized in that Set the loss function of the policy network to Where c is a constant, s represents the problem instance, which consists of the flow data from time point t-T+1 to time point t, and p φ (Ω|s) represents the probability of selecting the node set Ω under the problem instance s, φ is the parameter of the policy network, log is the logarithmic probability, Represents the gradient.

Images