Subgraph-driven lightweight federated learning method for spatiotemporal cellular network prediction

Abstract

The present invention belongs to the field of cellular network communication technology, and specifically relates to a subgraph-driven lightweight federated learning spatiotemporal cellular network prediction method. The method comprises the following steps: S1, data preprocessing; S2, subgraph partitioning; S3, model training; S4, model parameter pruning; S5, model transmission; S6, parameter aggregation; and S7, model update. A distributed subgraph partitioning strategy centered on the spatiotemporal correlation between nodes is designed for federated distributed environments. Based on this, a subgraph node partitioning scheme closer to actual FL application scenarios is developed. The present invention integrates the FL framework with the STGCN model and introduces an adaptive pruning scheme based on transmission parameter weights. During the model data transmission phase, a reinforcement learning agent is designed to make decisions and prune low-weight parameters to reduce model communication and computational costs, achieving model lightweighting while effectively maintaining prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of cellular network communication technology, and specifically relates to a subgraph-driven lightweight federated learning spatiotemporal cellular network prediction method. Background Art

[0002] With the advancement of 5G and the upcoming 6G technologies, mobile cellular network traffic forecasting has become a core element in modern and future cellular network management. The deployment of advanced network technologies has triggered the need for enhanced network capacity and more refined resource allocation mechanisms to meet the growing demand for data services. In particular, the demand for heterogeneous services is gradually expanding in applications that rely heavily on network reliability and low latency, such as autonomous driving and augmented / virtual reality. As a key component of telecommunications network management, cellular traffic forecasting will be crucial for achieving more intelligent distributed wireless network management during the future rollout of 6G technology. Powered by artificial intelligence solutions, advanced network traffic forecasting strategies can help network operators make more informed decisions in various areas, including network planning, traffic scheduling, fault diagnosis, and operational cost control. While research in cellular network traffic forecasting has traditionally relied on centralized learning strategies, the technical paradigm has evolved from traditional machine learning methods to deep learning solutions, with the recent introduction of advanced models based on graph neural networks.

[0003] However, with increasing demands for privacy protection and decentralized data management, the application of the federated learning (FL) framework in cellular network traffic forecasting is gaining increasing attention. Clients such as roadside units (RSUs), base stations (BSs), or institutions can collaboratively train models under the coordination of a central server, protecting the privacy of local data while also realizing the benefits of resource integration. With advancements in communication technology and device computing power, next-generation edge communication technologies, supported by 6G networks and beyond, are making the implementation and viability of federated learning a reality.

[0004] Despite this, the application of federated learning in the field of cellular traffic prediction faces multiple challenges:

[0005] (1) Challenges of spatiotemporal correlation and distributed scenarios: Cellular network base stations are spread across every corner of the city. Factors such as user mobility and dynamic changes in demand add complexity to cellular traffic prediction. Traffic patterns show diverse spatiotemporal correlations between different base stations, which are far beyond the reach of simple spatial adjacency or similar traffic pattern analysis methods. High-quality spatiotemporal correlations can help models achieve higher prediction accuracy. In a distributed federated learning environment, the fact that data cannot leave the local client provides data privacy protection for customers, but at the same time brings greater challenges to achieving high-precision predictions. This requires higher-quality technologies to maximize the extraction and preservation of node spatiotemporal properties to achieve accurate predictions.

[0006] (2) Considering communication and training costs: In a federated learning architecture, each client is usually an edge device, base station, or paid data provider with a certain level of computing power. During the training process, the client needs to frequently interact with the central server to aggregate and distribute model parameters. The amount of communication data and communication costs become factors that must be considered. The computing and storage resources of these clients are relatively limited, and the communication bandwidth is usually restricted. How to reduce training and data transmission costs while maintaining prediction accuracy has become an important consideration when designing a federated learning framework.

[0007] This paper discloses a subgraph-driven, lightweight federated learning method for spatiotemporal cellular network prediction. It designs a distributed subgraph partitioning strategy for federated distributed environments, centered on spatiotemporal correlations between nodes. This strategy extracts adjacency distances between nodes and captures traffic patterns to construct a sparse mixing matrix. From this matrix, low-dimensional continuous vectors representing the multi-hop neighborhood structure of nodes are extracted to represent the complex dependencies between nodes. This approach develops a subgraph node partitioning scheme that is more closely aligned with practical FL application scenarios. Furthermore, the FL framework is integrated with the STGCN (Spatiotemporal Graph Convolutional Neural Network) model, introducing an adaptive pruning scheme based on transmission parameter weights. During the model data transmission phase, a reinforcement learning agent is designed to make decisions and prune low-weight parameters to reduce model communication and computational costs, achieving model lightweighting while effectively maintaining prediction accuracy. Summary of the Invention

[0008] This paper, targeting cellular network scenarios, provides a lightweight spatiotemporal network traffic prediction method within a subgraph-driven federated learning framework. This method first employs a subgraph partitioning strategy to address the difficulty in extracting and retaining spatiotemporal correlations in the distributed cellular network environment, thereby improving model training accuracy. Furthermore, by combining federated learning with spatiotemporal graph convolutional neural networks, an adaptive pruning scheme based on transmission parameter weights is introduced to lightweight the training process, reducing transmission and computational costs in practical applications. Furthermore, through the designed reinforcement learning agent, the model's adaptive pruning decision process is optimized, achieving a training effect that reduces transmission costs while improving model accuracy.

[0009] The technical solutions adopted by the present invention are as follows:

[0010] The subgraph-driven lightweight federated learning spatiotemporal cellular network prediction method includes the following steps:

[0011] S1. Data preprocessing: Select the cellular network dataset and clean it to remove invalid or missing data points;

[0012] The data is processed into time series input, each sequence contains 12 time steps of historical data, which is used to predict the traffic situation in the next time step;

[0013] S2. Subgraph partitioning: Extract the adjacency distance matrix between nodes and the node traffic pattern matrix to construct a hybrid sparse matrix. Through graph embedding, extract low-dimensional continuous vectors representing the multi-hop neighborhood structure of nodes to represent the complex dependencies between nodes, obtain a similarity matrix, and perform community detection on this similarity matrix to obtain the final subgraph.

[0014] S3. Model training: Perform local client spatiotemporal graph convolutional neural network model training on the divided subgraph to obtain local model parameters;

[0015] S4, model parameter pruning: adaptively prune the trained model weight parameters, adjust the pruning strategy through the reinforcement learning agent, and obtain the pruned weight parameters;

[0016] S5, model transmission: the client transmits the pruned model parameters to the server;

[0017] S6, parameter aggregation: The server obtains the parameters passed by each client and performs federated average aggregation;

[0018] S7, model update: The server sends the aggregated new parameters back to the client, and the client obtains the returned parameters for the next round of spatiotemporal graph convolutional neural network model training until the end of the iteration round.

[0019] The hybrid sparse matrix considers the daily traffic of the BS record node as a vector, then The traffic generated within a day can be divided into a set of vectors. For any node That is, there exists a vector group ; Consider the vector group as a probability distribution, and each vector as a probability mass of the probability distribution, and get the node The probability mass Probability distribution of components :

[0020] The conversion cost between probability distributions of different nodes can be regarded as the distance between the traffic patterns of nodes. The Wasserstein probability distribution distance is used to calculate the traffic pattern distance between each two nodes, and the cost function It can be expressed as the difference between 1 and the cosine similarity between two vector groups:

[0021] Further get the node With node The Wasserstein probability distribution distance matrix between , the probability distribution distance between each two nodes for:

[0022] in, is a joint probability distribution that satisfies the following marginal constraints:

[0023] in, and Node and nodes The probability distribution of

[0024] Traffic pattern relationship matrix between nodes represented by probability distribution distance and the adjacency matrix dominated by the actual distances between nodes It can be understood as the external manifestation of implicit and explicit spatiotemporal relationships. The organic combination of the two can help nodes obtain more relationships.

[0025] Adjacency Matrix and Perform aggregation to obtain a new sparse matrix:

[0026] in, is the normalization constant, Represents the sparsity matrix preservation threshold.

[0027] The similarity matrix means that each node in the global graph has many neighboring nodes in both breadth and depth dimensions. It is necessary to further extract the information of these neighboring nodes to ensure the maximum effectiveness of the information contained in the subgraph after node division;

[0028] Use the Node2vec node embedding model to extract two features between nodes:

[0029] Homogeneity, that is, the closer the nodes are on the graph, the more similar they are;

[0030] Isomorphism, that is, if the structures around two nodes are similar, they are more similar;

[0031] By extracting homogeneity and isomorphism, the low-dimensional node continuous vector contains the neighbor relationship between multi-hop nodes, which can help further obtain the influence relationship between nodes; control the return parameters and in and out parameters To balance the DFS and BFS strategies, thus generating a node sequence, the node continuous vector extraction process is as follows:

[0032] Given the current node and the previous node , select the next node The transition probability for:

[0033] in, is the weight, is the normalization constant, weight The calculation depends on and :

[0034] Afterwards, for each node conduct The secondary length is The random walk generates node sequences and uses the generated node sequences to train the Skip-gram model. The Skip-gram model maximizes the node and context nodes The co-occurrence probability is used to learn the embedding representation of the node:

[0035] Among them, for the context node ,in Within the window size; is the set of all generated node pairs, is the conditional probability, expressed using the softmax function:

[0036] The final output of the model is the node embedding matrix , whose size is ,in is the number of nodes, is the dimension of the embedding vector, each row corresponds to the embedding vector of a node ;

[0037] In the global graph, there are nodes, and the embedding vector dimension of each node is , then the node embedding matrix It can be expressed as: in, is a node Embedding vector of

[0038] These extracted continuous vectors effectively represent the connection relationship and edge weight of each node within a certain dimension. Through algorithm calculation, the vector of each node can be obtained. , for every two nodes in the graph Vector similarity calculation can be performed to obtain the node's neighbor similarity matrix ,as follows:

[0039] .

[0040] The community detection partitioning refers to the use of similarity matrix , the Leiden algorithm is applied for community detection. The Leiden algorithm maximizes the modularity To optimize community structure and modularity Defined as:

[0041] in, is the adjacency matrix, and Node and nodes The degree, is the total number of edges in the graph, is the Kronecker delta, when the node and nodes When they belong to the same community, the value is 1;

[0042] or node and nodes If they do not belong to the same community, the value is 0;

[0043] The Leiden algorithm iteratively optimizes modularity and continuously refines the community structure to generate densely connected subgraphs.

[0044] In step S3, a network traffic prediction model based on a spatiotemporal graph convolutional neural network in an acyclic network is built, which uses federated learning as a communication framework, an adaptive pruning module as a transmission cost optimization module, and a local training model considering the limited client training resources.

[0045] The spatiotemporal graph convolutional neural network includes a spatiotemporal graph convolution block and a final convolution layer. Each spatiotemporal graph convolution block includes a Chebyshev graph convolution layer, a temporal convolution layer and a residual connection layer, where the input signal The shape is ,in, is the batch size, is the number of nodes, is the input feature dimension, is the time step, and the spatiotemporal graph convolution block is composed as follows:

[0046] Chebyshev graph convolutional layer:

[0047] The Chebyshev polynomial approximation method is used to calculate graph convolution. The calculation formula of the graph convolution layer is:

[0048] in, It is Chebyshev polynomials of order, is the normalized Laplace operator, It is The filter parameters of the order are output The shape is ,in is the output feature dimension;

[0049] Output of the temporal convolution layer to the graph convolution layer Perform convolution operation, the calculation formula is:

[0050] Output The shape is ,in is the time step after convolution;

[0051] In the residual connection layer, the input signal And the signal after graph convolution and time convolution Add to retain the information of the input signal; the calculation formula of the residual connection layer is:

[0052] in, Is the input signal The convolution operation is performed to make its shape Consistent; the shape of the residual connection layer output is ;

[0053] The final convolution layer further processes the output of the residual connection layer to obtain the final prediction result. The output shape of the final convolution layer is ,in is the final time step.

[0054] The federated learning communication framework includes a central server and an edge client;

[0055] After the model is initialized, it is distributed to each client. After training on the client, the model parameters are transmitted to the central server for aggregation, and then the parameters are distributed until the model converges.

[0056] Before each model aggregation, an adaptive weight pruning module is set up in the client. The adaptive module has a built-in reinforcement learning model and a dynamic loss window. The model uses the lowest training loss and validation loss within the local client's current model and the sliding window round to adaptively adjust the pruning ratio through the reinforcement learning model to protect model accuracy.

[0057] The federated learning communication framework is as follows:

[0058] Assume that there is clients, each client have data samples, the total amount of data is , each client trains a local model and updates its parameters;

[0059] Then, these parameters are uploaded to the central server, which performs a weighted average of the parameters based on the amount of data from each client and updates the global model as follows:

[0060] Model parameters for each client According to the amount of data Weighted to get the global model parameters .

[0061] In step S4, in order to further reduce the computation and communication costs, an adaptive parameter pruning module is introduced;

[0062] This module dynamically adjusts the pruning ratio according to the changes in the loss function during model training, thereby minimizing transmission and computing overhead while ensuring model accuracy. The details are as follows:

[0063] Assume that, in During the training round, each client The local model parameters are , the pruning ratio is , the model parameters after pruning are ; Adaptive pruning is performed through the following steps:

[0064] M1: Calculate the number of weights that need to be retained as follows:

[0065] in, is the total weight of the model, is the number of weights that need to be retained;

[0066] M2: Sort by the absolute value of the weight, as follows:

[0067] Model parameters Sort by the absolute value of The weight value is used as the threshold ;

[0068] M3: Apply threshold pruning as follows:

[0069] in, Is an indicator function, when the absolute value of the weight is greater than or equal to the threshold The value is 1 when , otherwise it is 0;

[0070] After each round of training, the minimum loss function memorized by the active loss window is compared with the loss function of the current round. The difference between and pruning interval The reinforcement learning agent model is input together to adaptively adjust the pruning ratio.

[0071] The reinforcement learning agent selects a Q-learning model, which can adaptively adjust actions to obtain the optimal solution based on the relationship between rewards and states. The model update algorithm is as follows:

[0072] in, Indicates the current state, Indicates the current action. is the learning rate, It is the reward of the next moment, is the discount factor.

[0073] The specific adjustment strategy of the adaptive adjustment action is as follows:

[0074] Action 1: linearly increase the pruning rate and Action 2: exponentially decrease the pruning rate, as follows:

[0075] Action 3: Increase the pruning interval and fine-tune the pruning rate as follows:

[0076] Action 4: Reduce the pruning interval and fine-tune the pruning rate as follows:

[0077] Action 5: Increase the pruning rate linearly by a small amount and Action 6: Decrease the pruning rate linearly by a small amount, as follows:

[0078] .

[0079] The technical effects achieved by the present invention are:

[0080] Compared to currently used partitioning methods based on adjacency matrices or relying solely on traffic patterns, the proposed subgraph-driven lightweight federated learning spatiotemporal cellular network prediction method more precisely handles the complexity of spatiotemporal relationships in federated distributed environments. This strategy uses a sparse hybrid relationship matrix, combined with embedding extraction of homogeneity and isomorphism relationships between nodes and community detection technology, to effectively preserve the complex connections between client nodes. This method maintains a high degree of correlation within each client's subgraph, significantly improving the accuracy of traffic prediction.

[0081] The subgraph-driven lightweight federated learning spatiotemporal cellular network prediction method introduced in this paper utilizes an adaptive parameter pruning strategy that, compared to other prediction models, prioritizes the data transmission costs faced by models in a federated environment. This strategy also integrates a reinforcement learning agent to dynamically adjust pruning ratios and parameter updates in real time, significantly reducing transmission costs while maintaining or even improving model prediction accuracy. Comparisons with other models have demonstrated that this method outperforms existing technologies in capturing spatiotemporal data characteristics, improving prediction accuracy, and addressing transmission costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 It is a flowchart of the subgraph-driven lightweight federated learning spatiotemporal cellular network prediction method of the present invention;

[0083] Figure 2 It is an architectural diagram of the centralized prediction model and the federated distributed prediction model of the present invention;

[0084] Figure 3 This is a structural diagram of the overall model framework of the subgraph-driven lightweight federated learning spatiotemporal cellular network prediction method of the present invention;

[0085] Figure 4 It is a histogram comparing the evaluation indicators MAE and RMSE of the subgraph partitioning strategy of the present invention and the baseline method;

[0086] Figure 5 It is a histogram comparing the evaluation indicators MAE and RMSE of SMS in the Telecom Italia dataset of the present invention;

[0087] Figure 6 It is a histogram comparing the evaluation indicators MAE and RMSE of calls in the Telecom Italia dataset of the present invention;

[0088] Figure 7 It is a histogram comparing the evaluation indicators MAE and RMSE of the flow in the Telecom Italia dataset of the present invention;

[0089] Figure 8 It is a histogram comparing the evaluation indicators of the flow data on the Pems07 dataset of the present invention;

[0090] Figure 9 It is a line chart showing the comparison of the prediction of real call data based on the Telecom Italia dataset of the present invention. DETAILED DESCRIPTION

[0091] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is described in detail below with reference to the following examples. It should be understood that the following text is only used to describe one or more specific embodiments of the present invention and does not strictly limit the scope of protection of the present invention.

[0092] like Figure 1-9 As shown in the figure, the subgraph-driven lightweight federated learning spatiotemporal cellular network prediction method, including the subgraph partitioning strategy and adaptive parameter pruning scheme, is as follows:

[0093] First, the subgraph partitioning strategy involves creating a hybrid sparse matrix that captures explicit and implicit spatiotemporal relationships between nodes, which is crucial for effectively distributing workloads among clients in a federated learning environment while retaining the necessary relational information for accurate predictions, as follows:

[0094] (a) Hybrid sparse matrix;

[0095] Preprocess the data to obtain a hybrid relationship matrix, and regard the daily traffic of the BS record node as a vector. The traffic generated within a day can be divided into a set of vectors. For any node That is, there exists a vector group ; Consider the vector group as a probability distribution, and each vector as a probability mass of the probability distribution, and get the node The probability mass Probability distribution of components :

[0096] The conversion cost between probability distributions of different nodes can be regarded as the distance between the traffic patterns of nodes. The Wasserstein probability distribution distance is used to calculate the traffic pattern distance between each two nodes, and the cost function It can be expressed as the difference between 1 and the cosine similarity between two vector groups:

[0097] Further get the node With node The Wasserstein probability distribution distance matrix between , the probability distribution distance between each two nodes for:

[0098] in, is a joint probability distribution that satisfies the following marginal constraints:

[0099] in, and Node and nodes The probability distribution of

[0100] Traffic pattern relationship matrix between nodes represented by probability distribution distance and the adjacency matrix dominated by the actual distances between nodes It can be understood as the external manifestation of implicit and explicit spatiotemporal relationships. The organic combination of the two can help nodes obtain more relationships.

[0101] Adjacency Matrix and Perform aggregation to obtain a new sparse matrix:

[0102] in, is the normalization constant, represents the sparsity matrix retention threshold;

[0103] (b) extract low-dimensional information and obtain the adjacency similarity matrix;

[0104] Each node in the global graph has many neighboring nodes in both breadth and depth. It is necessary to further extract the information of these neighboring nodes to ensure the maximum effectiveness of the information contained in the subgraph after node partitioning.

[0105] Using the Node2vec node embedding model, we focus on extracting two features between nodes:

[0106] (A) Homogeneity, that is, the closer the distance between nodes on the graph, the more similar they are;

[0107] (B) Isomorphism, that is, if the structures around two nodes are similar, they are more similar;

[0108] By extracting homogeneity and isomorphism, the low-dimensional node continuous vector contains the neighbor relationship between multi-hop nodes, which can help further obtain the influence relationship between nodes; control the return parameters and in and out parameters To balance the DFS (depth-first search) and BFS (breadth-first search) strategies, thus generating a node sequence, the node continuous vector extraction process is as follows:

[0109] Given the current node and the previous node , select the next node The transition probability for:

[0110] in, is the weight, is the normalization constant, weight The calculation depends on and :

[0111] Afterwards, for each node conduct The secondary length is The random walk generates node sequences and uses the generated node sequences to train the Skip-gram model. The Skip-gram model maximizes the node and context nodes The co-occurrence probability is used to learn the embedding representation of the node:

[0112] Among them, for the context node ;in Within the window size range, is the set of all generated node pairs, is the conditional probability, expressed as a softmax function:

[0113] The final output of the model is the node embedding matrix , whose size is ,in is the number of nodes, is the dimension of the embedding vector, each row corresponds to the embedding vector of a node ;

[0114] In the global graph, there are nodes, and the embedding vector dimension of each node is , then the node embedding matrix It can be expressed as:

[0115] in, is a node Embedding vector of

[0116] These extracted continuous vectors effectively represent the connection relationship and edge weight of each node within a certain dimension. Through algorithm calculation, the vector of each node can be obtained. , for every two nodes in the graph Vector similarity calculation can be performed to obtain the node's neighbor similarity matrix ,as follows:

[0117] (c) Community detection node division to obtain a subgraph with higher density;

[0118] Using the similarity matrix , the Leiden algorithm is applied for community detection. The Leiden algorithm maximizes the modularity To optimize community structure and modularity Defined as:

[0119] in, is the adjacency matrix, and Node and nodes The degree, is the total number of edges in the graph, is the Kronecker delta (Kronecker function), when the node and nodes If they belong to the same community, the value is 1, otherwise it is 0;

[0120] The Leiden algorithm iteratively optimizes modularity and continuously refines the community structure, thereby generating densely connected subgraphs;

[0121] Based on the divided communities, obtain the subgraph ,in is the desired number of clients; each subgraph Contains more densely connected nodes;

[0122] Traditional federated learning prediction frameworks typically use simple linear neural network models such as ANN or LSTM. However, with technological advancements and increasing prediction accuracy requirements, more complex models have emerged. Graph neural network models are based on the graph structure between nodes, which is the foundation for model training. In a distributed architecture, this is reflected in the importance of partitioning subgraph sets, as discussed previously. While providing better performance, models based on graph neural networks emphasize the importance of device computing power.

[0123] To achieve better performance while minimizing computational and transmission costs, we introduced an adaptive fine-grained pruning module that maintains the granularity of pruned data at the parameter level. Pruned parameters not only reduce the cost of transmitting data between the client and server, but also, due to the parametric nature of the model, effectively reduce model training complexity and thus shorten training time. An adaptive mechanism based on reinforcement learning agents enables the model to achieve these advantages while ensuring training accuracy.

[0124] Specifically, a network traffic prediction model was built using federated learning as the communication framework, an adaptive pruning module to optimize transmission costs, and a local training model that uses a spatiotemporal graph convolutional neural network in an acyclic network to consider the limited client training resources.

[0125] The spatiotemporal graph convolutional neural network consists of a spatiotemporal graph convolution block and a final convolution layer. Each spatiotemporal graph convolution block consists of a Chebyshev graph convolution layer, a temporal convolution layer, and a residual connection layer, where the input signal The shape is ,in, is the batch size, is the number of nodes, is the input feature dimension, is the time step, and the spatiotemporal graph convolution block is composed as follows:

[0126] Chebyshev graph convolutional layer:

[0127] The Chebyshev polynomial approximation method is used to calculate graph convolution. The calculation formula of the graph convolution layer is:

[0128] in, It is Chebyshev polynomials of order, is the normalized Laplace operator, It is The filter parameters of the order are output The shape is ,in is the output feature dimension;

[0129] Output of the temporal convolution layer to the graph convolution layer Perform convolution operation, the calculation formula is:

[0130] Output The shape is ,in is the time step after convolution;

[0131] In the residual connection layer, the input signal And the signal after graph convolution and time convolution Add to retain the information of the input signal; the calculation formula of the residual connection layer is:

[0132] in, Is the input signal The convolution operation is performed to make its shape Consistent; the shape of the residual connection layer output is ;

[0133] The final convolution layer further processes the output of the residual connection layer to obtain the final prediction result. The output shape of the final convolution layer is ,in is the final time step;

[0134] A federated learning communication framework with a central server and edge clients was built. After model initialization, the model was distributed to each client. After training on the client, the model parameters were transmitted to the central server for aggregation, and then the parameters were distributed until the model converged. Before each model aggregation, an adaptive weight pruning module was set up in the client. The adaptive module has a built-in reinforcement learning model and a dynamic loss window. The model uses the lowest training loss and validation loss within the current round of the local client model and the sliding window round to adaptively adjust the pruning ratio through the reinforcement learning model to protect the model accuracy. The pruned weights of each client are uploaded to the central server to achieve the purpose of reducing model transmission costs. Since pruning reduces computational complexity to a certain extent, the model computation cost is also reduced.

[0135] The federated learning communication framework is as follows: Assume that clients, each client have data samples, the total amount of data is , each client trains a local model and updates its parameters, and then uploads these parameters to the central server. The server performs a weighted average of the parameters based on the amount of data from each client and updates the global model as follows:

[0136] Model parameters for each client According to the amount of data Weighted to get the global model parameters ;

[0137] Second, to further reduce computational and communication costs, an adaptive parameter pruning module is introduced. This module dynamically adjusts the pruning ratio based on changes in the loss function during model training, thereby minimizing transmission and computational overhead while ensuring model accuracy. The details are as follows:

[0138] Assume that, in During the training round, each client The local model parameters are , the pruning ratio is , the model parameters after pruning are ; Adaptive pruning is performed through the following steps:

[0139] 1. Calculate the number of weights that need to be retained:

[0140] in, is the total weight of the model, is the number of weights that need to be retained;

[0141] 2. Sort by the absolute value of the weight:

[0142] Model parameters Sort by the absolute value of The weight value is used as the threshold ;

[0143] 3. Apply threshold for pruning:

[0144] in, Is an indicator function, when the absolute value of the weight is greater than or equal to the threshold The value is 1 when , otherwise it is 0;

[0145] After each round of training, the minimum loss function memorized by the active loss window is compared with the loss function of the current round. The difference between and pruning interval The reinforcement learning agent model is input together to adaptively adjust the pruning ratio;

[0146] The reinforcement learning agent selects the Q-learning model, which can adaptively adjust actions to obtain the optimal solution based on the relationship between rewards and states. The model update algorithm is as follows:

[0147] in, Indicates the current state, Indicates the current action. is the learning rate, It is the reward of the next moment, is the discount factor;

[0148] The specific adjustment strategy of the action is as follows:

[0149] Action 1: linearly increase the pruning rate and Action 2: exponentially decrease the pruning rate, as follows:

[0150] Action 3: Increase the pruning interval and fine-tune the pruning rate as follows:

[0151] Action 4: Reduce the pruning interval and fine-tune the pruning rate as follows:

[0152] Action 5: Increase the pruning rate linearly by a small amount and Action 6: Decrease the pruning rate linearly by a small amount, as follows:

[0153] .

[0154] The foregoing is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained herein shall, unless otherwise specified or limited, be implemented in accordance with conventional means in the art.

Claims

1. A subgraph-driven lightweight federated learning spatiotemporal cellular network prediction method, characterized by: The following steps are involved: S1. Data preprocessing: Select the cellular network dataset and clean it to remove invalid or missing data points; The data is processed into time series input, each sequence contains 12 time steps of historical data, which is used to predict the traffic situation in the next time step; S2. Subgraph partitioning: Extract the adjacency distance matrix between nodes and the node traffic pattern matrix to construct a hybrid sparse matrix. Use graph embedding to extract low-dimensional continuous vectors representing the multi-hop neighborhood structure of nodes to represent the complex dependencies between nodes. Obtain a similarity matrix, and perform community detection on the similarity matrix to obtain the final subgraph. The hybrid sparse matrix regards the daily traffic of the BS record node as a vector, and the traffic generated within D days is divided into a group of vectors by day. That is, there exists a vector group ; Consider the vector group as a probability distribution, and each vector as a probability mass of the probability distribution, and get the node The probability mass Probability distribution of components : The conversion cost between the probability distributions of different nodes is regarded as the gap distance between the node traffic patterns. The Wasserstein probability distribution distance is used to calculate the traffic pattern distance between each two nodes, and the cost function Expressed as the difference between 1 and the cosine similarity between two vector groups: Further get the node With node The Wasserstein probability distribution distance matrix between , the probability distribution distance between each two nodes for: in, is a joint probability distribution that satisfies the following marginal constraints: in, and Node and nodes The probability distribution of Traffic pattern relationship matrix between nodes represented by probability distribution distance and the adjacency matrix dominated by the actual distances between nodes It is understood as the external manifestation of implicit and explicit space-time relations, and the organic combination of the two; Adjacency Matrix and Perform aggregation to obtain a new sparse matrix: in, is the normalization constant, represents the sparsity matrix retention threshold; The similarity matrix means that each node in the global graph has many neighboring nodes in both breadth and depth dimensions. It is necessary to further extract the information of these neighboring nodes to ensure the maximum effectiveness of the information contained in the subgraph after node division; Use the Node2vec node embedding model to extract two features between nodes: Homogeneity, where the closer the nodes are on the graph, the more similar they are; Isomorphism, which means that if the structures around two nodes are similar, they are more similar; By extracting homogeneity and isomorphism, the low-dimensional node continuous vector contains the neighbor relationship between multi-hop nodes, which helps to further obtain the influence relationship between nodes; control the return parameters and in and out parameters To balance the DFS and BFS strategies, thus generating a node sequence, the node continuous vector extraction process is as follows: Given the current node and the previous node , select the next node The transition probability for: in, is the weight, is the normalization constant, weight The calculation depends on and : Afterwards, for each node conduct The secondary length is The random walk generates node sequences and uses the generated node sequences to train the Skip-gram model. The Skip-gram model maximizes the node and context nodes The co-occurrence probability is used to learn the embedding representation of the node: Among them, for the context node , Within the window size; is the set of all generated node pairs, is the conditional probability, expressed as a softmax function: The final output of the model is the node embedding matrix , whose size is ,in is the number of nodes, is the dimension of the embedding vector, each row corresponds to the embedding vector of a node ; In the global graph, there are nodes, and the embedding vector dimension of each node is , then the node embedding matrix Expressed as: in, is a node Embedding vector of These extracted continuous vectors effectively represent the connection relationship and edge weight of each node within a certain dimension. Through algorithm calculation, the vector of each node is obtained. , for every two nodes in the graph Vector similarity calculation is performed to obtain the node's neighbor similarity matrix ,as follows: The community detection partitioning refers to the use of similarity matrix , the Leiden algorithm is applied for community detection. The Leiden algorithm maximizes the modularity To optimize community structure and modularity Defined as: in, is the adjacency matrix, and Node and nodes The degree, is the total number of edges in the graph, is the Kronecker delta, when the node and nodes When they belong to the same community, the value is 1; or and nodes When they do not belong to the same community, the value is 0; the Leiden algorithm iteratively optimizes the modularity and continuously refines the community structure, thereby generating a densely connected subgraph; S3. Model training: Perform local client spatiotemporal graph convolutional neural network model training on the divided subgraph to obtain local model parameters; In step S3, a federated learning communication framework is also built. The adaptive pruning module is a module for optimizing transmission costs. The local training model takes into account the limited training resources of the client, so a network traffic prediction model of the spatiotemporal graph convolutional neural network in the acyclic network is adopted. The spatiotemporal graph convolutional neural network includes a spatiotemporal graph convolution block and a final convolution layer. Each spatiotemporal graph convolution block includes a Chebyshev graph convolution layer, a temporal convolution layer and a residual connection layer, where the input signal The shape is ,in, is the batch size, is the number of nodes, is the input feature dimension, is the time step, and the spatiotemporal graph convolution block is composed as follows: Chebyshev graph convolutional layer: The Chebyshev polynomial approximation method is used to calculate graph convolution. The calculation formula of the graph convolution layer is: in, It is Chebyshev polynomials of order, is the normalized Laplace operator, It is The filter parameters of the order are output The shape is ,in is the output feature dimension; Output of the temporal convolution layer to the graph convolution layer Perform convolution operation, the calculation formula is: Output The shape is ,in is the time step after convolution; In the residual connection layer, the input signal And the signal after graph convolution and time convolution Add to retain the information of the input signal; the calculation formula of the residual connection layer is: in, Is the input signal The convolution operation is performed to make its shape Consistent; the shape of the residual connection layer output is ; The final convolution layer further processes the output of the residual connection layer to obtain the final prediction result. The output shape of the final convolution layer is ,in is the final time step; The federated learning communication framework includes a central server and an edge client; After the model is initialized, it is distributed to each client. After training on the client, the model parameters are transmitted to the central server for aggregation, and then the parameters are distributed until the model converges. Before each model aggregation, an adaptive weight pruning module is set up in the client. The adaptive module has a built-in reinforcement learning model and a dynamic loss window. The model uses the lowest training loss and validation loss within the current round of the local client model and the sliding window round to adaptively adjust the pruning ratio through the reinforcement learning model to protect model accuracy. The pruned weights of each client are uploaded to the central server; the federated learning communication framework is as follows: Assume that there is clients, each client have data samples, the total amount of data is , each client trains a local model and updates its parameters; Then, these parameters are uploaded to the central server, which performs a weighted average of the parameters based on the amount of data from each client and updates the global model as follows: Model parameters for each client According to the amount of data Weighted to get the global model parameters ; S4, model parameter pruning: adaptively prune the trained model weight parameters, adjust the pruning strategy through the reinforcement learning agent, and obtain the pruned weight parameters; S5, model transmission: the client transmits the pruned model parameters to the server; S6, parameter aggregation: The server obtains the parameters passed by each client and performs federated average aggregation; S7, model update: The server sends the aggregated new parameters back to the client, and the client obtains the returned parameters for the next round of spatiotemporal graph convolutional neural network model training until the end of the iteration round.

2. The subgraph-driven lightweight federated learning spatiotemporal cellular network prediction method according to claim 1, characterized in that: In step S4, in order to further reduce the computation and communication costs, an adaptive parameter pruning module is introduced; This module dynamically adjusts the pruning ratio according to the changes in the loss function during model training, thereby minimizing transmission and computing overhead while ensuring model accuracy. The details are as follows: Assume that, in During the training round, each client The local model parameters are , the pruning ratio is , the model parameters after pruning are ; Adaptive pruning is performed through the following steps: M1: Calculate the number of weights that need to be retained as follows: in, is the total weight of the model, is the number of weights that need to be retained; M2: Sort by the absolute value of the weight, as follows: Model parameters Sort by the absolute value of The weight value is used as the threshold ; M3: Apply threshold pruning as follows: in, Is an indicator function, when the absolute value of the weight is greater than or equal to the threshold The value is 1 when , otherwise it is 0; After each round of training, the minimum loss function memorized by the active loss window is compared with the loss function of the current round. The difference between and pruning interval The reinforcement learning agent model is input together to adaptively adjust the pruning ratio.

3. The subgraph-driven lightweight federated learning spatiotemporal cellular network prediction method according to claim 1, characterized in that: The reinforcement learning agent model selects the Q-learning model. The Q-learning model adaptively adjusts the action to obtain the optimal solution based on the relationship between reward and state. The model update algorithm is as follows: in, Indicates the current state, Indicates the current action. is the learning rate, It is the reward of the next moment, is the discount factor.

4. The subgraph-driven lightweight federated learning spatiotemporal cellular network prediction method according to claim 3, characterized in that: The specific adjustment strategy of the adaptive adjustment action is as follows: Action 1: linearly increase the pruning rate and Action 2: exponentially decrease the pruning rate, as follows: Action 3: Increase the pruning interval and fine-tune the pruning rate as follows: Action 4: Reduce the pruning interval and fine-tune the pruning rate as follows: Action 5: Increase the pruning rate linearly by a small amount and Action 6: Decrease the pruning rate linearly by a small amount, as follows: 。

Images