Online computing unloading method and system for edge computing device of graph neural network
By constructing a node feature matrix for edge computing devices using a graph convolutional neural network, generating offloading probabilities and updating the model, the challenges of computation offloading decisions and resource allocation in traditional mobile edge computing networks are solved, achieving low-latency, high-efficiency computation offloading and device utilization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NARI TECH CO LTD
- Filing Date
- 2025-12-12
- Publication Date
- 2026-05-05
AI Technical Summary
Traditional mobile edge computing networks rely on deep neural networks (DNNs), which struggle to capture spatial relationships between users and fail to fully utilize topological connections between users. In dynamic multi-user environments, efficiently making computation offloading decisions and allocating resources is a major challenge.
A graph convolutional neural network (GCN) is used to construct the node feature matrix of the edge computing device. Feature extraction is performed through a two-layer GCN structure to generate the uninstallation probability. The uninstallation probability of each user is output through the sigmoid function. The model is trained and updated by combining experience replay to generate the optimal uninstallation decision.
It effectively alleviates the latency and energy consumption problems caused by cloud processing, realizes the efficient utilization of various edge computing devices in the power grid field, and has strong generalization performance.
Smart Images
Figure CN121979657A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of edge computing technology, and in particular to an online computation offloading method and system for edge computing devices using graph neural networks. Background Technology
[0002] As an emerging paradigm, edge computing pushes computing and storage capabilities to the network edge, utilizing various heterogeneous network devices to integrate computing, storage, and transmission resources in the edge network. This enables distributed deployment of application services and multi-point collaborative offloading of computing tasks, providing convenient low-latency, high-bandwidth intelligent services for mobile users and portable devices closer to the network edge.
[0003] In the power grid sector, with the rapid development of technologies such as the Internet of Things (IoT), artificial intelligence (AI), and big data, edge computing is gradually becoming a crucial computing architecture supporting these emerging technologies. Traditional cloud computing architectures, due to data transmission latency and bandwidth limitations, struggle to meet the demands for real-time processing and low-latency response. This is especially true in scenarios with high real-time and localized processing requirements, such as smart grid monitoring and management, distributed energy management, and electric vehicle charging management, where the importance of edge computing is increasingly evident. Simultaneously, edge computing can process sensitive data locally, reducing security risks during data transmission and further enhancing data privacy and security. With the accelerating pace of digital transformation, edge computing has become a vital component of next-generation information infrastructure.
[0004] With the increasing demand for low-latency and high-reliability computing from IoT terminal devices, mobile edge computing effectively alleviates the latency and energy consumption problems caused by cloud processing by offloading computing tasks to the network edge. Traditional mobile edge computing networks rely on deep neural networks (DNNs) to generate offloading decisions, but DNNs have difficulty capturing the spatial relationships between users and do not fully utilize the topological connections between users. In a dynamic multi-user environment, how to efficiently make computing offloading decisions and allocate resources is also a major challenge. Summary of the Invention
[0005] Purpose of the invention: The purpose of this invention is to provide an online computation offloading method and system for edge computing devices that help improve information processing efficiency and optimize device performance using graph neural networks.
[0006] Technical solution: The online computation offloading method for edge computing devices using graph neural networks described in this invention includes the following steps:
[0007] (1) Preprocess the input data to construct the node feature matrix of the edge computing device;
[0008] (2) Construct a GCN model, perform convolution processing on the graph matrix feature data to generate unloading probabilities. The GCN model uses a two-layer GCN structure for feature extraction. The first layer maps the node features to 64 dimensions, and the second layer further maps the features to 32 dimensions, using ReLU as the activation function.
[0009]
[0010]
[0011] in, For the input feature matrix, and These are the trainable weight matrices for the first and second layers, respectively;
[0012] Finally, a fully connected layer is used in the output layer to map the 32-dimensional features to 1-dimensional features, and the sigmoid function is used to output the uninstallation probability for each user.
[0013]
[0014] in, Given the output layer weight matrix, the Sigmoid activation function formula is:
[0015] ;
[0016] (3) Generate based on unloading probability The system identifies several candidate adaptive unloading decisions, evaluates the weighted computation rate of each strategy, and selects the optimal unloading decision.
[0017] (4) Based on the generated optimal unloading decision, the GCN model is updated by training through experience replay.
[0018] Further, in step (1), the input data is preprocessed to construct the node feature matrix of the edge computing device, including:
[0019] Based on the number of edge computing devices Set their respective channel gain values. Based on Gaussian similarity, the association information between users is calculated, and an adjacency matrix is generated.
[0020] Furthermore, the adjacency matrix formula is as follows:
[0021]
[0022] in, The symmetric normalized adjacency matrix is calculated as follows:
[0023]
[0024] in, Adjacency matrix The degree matrix.
[0025] Further, in step (3), K candidate adaptive unloading decisions are generated based on the unloading probability, where K is an adjustable parameter with a maximum of N+1, and the weighted computation rate of each strategy is evaluated to select the optimal unloading decision, including:
[0026] The first decision is obtained directly through binarization using a set threshold: values greater than the threshold are offloaded to the edge server, while values less than the threshold are computed locally.
[0027] The remaining K-1 actions are to sequentially flip the decision of the least uncertain device based on the order in which the output value of each device is close to the threshold.
[0028] Furthermore, the decision-making steps for the K candidate adaptive unloading are as follows:
[0029] (1) Calculate the absolute distance between the unloading probability value of each device and the set threshold;
[0030] (2) Sort the devices in ascending order based on their absolute distance, and put the least uncertain devices at the top to obtain the index list;
[0031] (3) For the first K-1 most uncertain users, reverse the user's decision in turn to generate new candidate decisions;
[0032] After obtaining K candidate decisions, calculate their corresponding weighted computation rate, and finally select the decision that brings the highest computation rate as the unloading decision for the current time frame through a binary evaluation method.
[0033] Furthermore, in step (3), the weighted calculation rate is:
[0034] The total system time is divided into several consecutive time frames of length T. In a static IoT environment, T is typically on the order of a few seconds, a relatively short time. It is assumed that the frame length is less than the channel coherence time. Each time frame can be divided into two phases:
[0035] Phase 1: Downlink Wireless Power Transmission Phase: At the beginning of time frame T, the receiving point broadcasts radio frequency power for a duration of aT, where... This is the time allocation factor for energy transfer; the device harvests energy from the received radio frequency signal during this phase.
[0036] Phase Two: Uplink Task Execution Phase: The remaining time (1-a)T is used for task execution and follows the binary offload strategy. The device must select one of the following two modes ( ):
[0037] (a) Local computation (x=0): Computation is performed locally using the collected energy throughout the entire (1-a)T duration;
[0038] (b) Complete offloading (x=1): The entire task is offloaded to the receiving point; the device uses the collected energy to upload the task data to the receiving point for remote execution; the time required for the receiving point to calculate the task and download the results back to the device is negligible compared to T.
[0039] Let h represent the wireless channel gain between the receiving point and the device, and let the total energy collected by the device within one time frame be:
[0040]
[0041] in, For the efficiency of the energy harvesting circuit;
[0042] Assume the device's processor has a computing speed of When local computing mode is selected, the energy consumed to execute computing tasks is Where k is the effective capacitance coefficient related to the chip structure, taken as... The local computing speed is:
[0043]
[0044] Assume the transmission power of the device is Then there is According to Shannon's formula, the uplink data transmission rate, i.e., the offload rate, is:
[0045]
[0046] in, For unified communication bandwidth, This refers to the noise power of the receiver;
[0047] When the receiving point is connected to N devices, and N>1, the local computing rate is:
[0048]
[0049] in, For the wireless channel gain between different devices and receiving points, The effective capacitance coefficient for different devices;
[0050] When the complete uninstallation mode is selected, set Let be the time allocation factor for the unloading of the i-th device, then we have The uninstallation rate can be derived as follows:
[0051]
[0052] Assuming that within a time frame, only the wireless channel gain h... i It is time-varying, while other system parameters are fixed. Maximizing the weighted sum computation rate can be expressed as:
[0053]
[0054] Among them, w i As weight; This indicates the decision to unload.
[0055] Further, in step (4), the GCN model is updated through experience replay training based on the generated optimal unloading decision, including:
[0056] The model maintains a fixed-size experience replay storage area with a capacity of M; at each time frame, the model updates the current state decision pair. Stored in the buffer, where This refers to the channel gain between the corresponding device and the receiving point. The optimal decision is obtained through the binary search method; when the buffer is full, new experience will overwrite the oldest experience.
[0057] The model employs a periodic training strategy. Every T time frames, a small batch of experience is randomly sampled from the experience replay buffer for training. Random sampling reduces the correlation between training samples and accelerates the convergence speed. The training process uses a binary cross-entropy loss function.
[0058] Furthermore, an online computation offloading system for an edge computing device using a graph neural network includes:
[0059] Data preprocessing module: preprocesses the input data and constructs the node feature matrix of the edge computing device;
[0060] Graph Convolution Module: Constructs a GCN model, performs convolution processing on graph matrix feature data, and generates unloading probabilities;
[0061] Unloading decision generation module: Generates K candidate adaptive unloading decisions based on the unloading probability, evaluates the weighted calculation rate of each strategy, and selects the optimal unloading decision;
[0062] Replay Training Module: Based on the generated optimal unloading decision, the GCN model is updated through experience replay training.
[0063] An online computation offloading system for edge computing devices using graph neural networks, comprising:
[0064] Data preprocessing module: preprocesses the input data and constructs the node feature matrix of the edge computing device;
[0065] Graph Convolution Module: Constructs a GCN model, performs convolution processing on graph matrix feature data, and generates unloading probabilities;
[0066] Unloading decision generation module: Generates K candidate adaptive unloading decisions based on the unloading probability, evaluates the weighted calculation rate of each strategy, and selects the optimal unloading decision;
[0067] Replay Training Module: Based on the generated optimal unloading decision, the GCN model is updated through experience replay training.
[0068] A computing device includes: one or more processors, one or more memories, and one or more programs, the programs being stored in the memories and configured to be executed by the processors, wherein the programs, when loaded onto the processors, implement the online computation offloading method for edge computing devices of the graph neural network.
[0069] A storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the online computation offloading method for an edge computing device of a graph neural network.
[0070] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. The present invention is based on graph convolutional neural network to implement a task offloading strategy for edge computing devices. By offloading computing tasks to the network edge, the latency and energy consumption problems caused by cloud processing are effectively alleviated; 2. It realizes the efficient utilization of various edge computing devices in the power grid field; 3. It has strong generalization performance. Attached Figure Description
[0071] Figure 1 This is a flowchart of the method of the present invention;
[0072] Figure 2 The figures show the ablation experiment results for different parameters M; among them, Figure 2 (a) is a comparison chart of different storage area sizes. Figure 2 (b) is a comparison chart of different learning rates. Figure 2 (c) is a comparison chart of the sizes of different batches. Figure 2 (d) is a comparison chart of different training intervals. Detailed Implementation
[0073] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0074] like Figure 1 As shown, the online computation offloading method for edge computing devices using graph neural networks includes:
[0075] Step 1: Preprocess the input data to construct the node feature matrix of the edge computing device;
[0076] Step 2: Construct a GCN model, perform convolution processing on the graph matrix feature data, and generate unloading probabilities. The GCN model employs a two-layer GCN structure for feature extraction. The first layer maps node features to 64 dimensions, and the second layer further maps the features to 32 dimensions, using ReLU as the activation function.
[0077]
[0078]
[0079] in, For the input feature matrix, and These are the trainable weight matrices for the first and second layers, respectively;
[0080] Finally, a fully connected layer is used in the output layer to map the 32-dimensional features to 1-dimensional features, and the sigmoid function is used to output the uninstallation probability for each user.
[0081]
[0082] in, Given the output layer weight matrix, the Sigmoid activation function formula is:
[0083] ;
[0084] Step 3: Generate based on uninstallation probability The system identifies several candidate adaptive unloading decisions, evaluates the weighted computation rate of each strategy, and selects the optimal unloading decision.
[0085] Step 4: Based on the generated optimal unloading decision, update the GCN model through experience replay training.
[0086] In step one, based on the number of edge computing devices Set their respective channel gain values. (dimension is) Based on Gaussian similarity, the association information between users is calculated, and an adjacency matrix is generated. The formula is:
[0087]
[0088] in, The standard deviation of the channel gain. The channel gain between the corresponding device and the receiving point is calculated; the symmetric normalized adjacency matrix is obtained:
[0089]
[0090] in, Adjacency matrix The degree matrix.
[0091] In step three, the uninstallation probability is generated. Several candidate adaptive unloading decisions are considered, and the weighted computational rate of each strategy is evaluated to select the optimal unloading decision, including:
[0092] (1) When m>0.5, unload the node and set m to 1; when m<0.5, do not unload the node and set it to 0.
[0093] (2) Calculate the difference between the probability m of the remaining nodes and 0.5, and find the K-1 nodes with the smallest difference (i.e., the nodes closest to 0.5). For the K-1 nodes, generate K-1 decisions and take different processing methods according to the node probabilities:
[0094] (a) If the node probability m[i] > 0.5, the node is set to 1. Nodes with a probability greater than the node probability are also set to 1, i.e., unloaded (i.e., nodes with m > m[i] are set to 1, and the rest are set to 0).
[0095] (b) If the node probability m[i] < 0.5, the node is set to 0. Only nodes with a probability greater than the node probability are set to 1, i.e., unloaded (i.e., nodes with m > m[i] are set to 1, and the rest are set to 0).
[0096] (3) Use the binary search algorithm to find the strategy whose probability m is closest to 0.5 among the weighted computation rates of each strategy. This strategy is the optimal decision.
[0097] (4) In order to reduce unnecessary candidate policy generation and improve the efficiency of calculating the optimal decision, this method adjusts the K value by recording the index of the best candidate policy selected in each time frame (i.e., which of the K policies is selected as the best);
[0098] According to recent The maximum index value appearing in a time frame Adjustment value:
[0099] .
[0100] In step three, the weighted calculation rate is:
[0101] The total system time is divided into several consecutive time frames of length T. In a static IoT environment, T is typically on the order of a few seconds, a relatively short time. It is assumed that the frame length is less than the channel coherence time. Each time frame can be divided into two phases:
[0102] Phase 1: Downlink Wireless Power Transmission Phase: At the beginning of time frame T, the receiving point broadcasts radio frequency power for a duration of aT, where... This is the time allocation factor for energy transfer; the device harvests energy from the received radio frequency signal during this phase.
[0103] Phase Two: Uplink Task Execution Phase: The remaining time (1-a)T is used for task execution and follows the binary offload strategy. The device must select one of the following two modes ( ):
[0104] (a) Local computation (x=0): Computation is performed locally using the collected energy throughout the entire (1-a)T duration;
[0105] (b) Complete unloading (x=1): The entire task is unloaded to the receiving point; the device uses the collected energy to upload the task data to the receiving point for remote execution; the time required for the receiving point to calculate the task and download the results back to the device is negligible compared to T.
[0106] Assume the uplink and downlink channels are reciprocal and remain constant within a time frame, but may vary between different time frames. Let h represent the wireless channel gain between the receiving point and the device, and the total energy collected by the device within a time frame is:
[0107]
[0108] in, The efficiency of the energy harvesting circuit.
[0109] Assume the device's processor has a computing speed of When local computing mode is selected, the energy consumed to execute computing tasks is Where k is the effective capacitance coefficient related to the chip structure, due to energy loss, To process more data under energy-constrained conditions, take The local computing speed is:
[0110]
[0111] When the device selects the fully unloaded mode, it is assumed that the receiving point's computing speed and transmitting power are much greater than those of the energy- and size-constrained device. Therefore, the time spent by the receiving point on task computing and downloading can be ignored. That is, the uplink task execution phase only includes uplink data transmission, and it will use all the collected energy E for uplink data transmission. Let the device's transmission power be... Then there is According to Shannon's formula, the uplink data transmission rate, i.e., the offload rate, is:
[0112]
[0113] in, For unified communication bandwidth, This represents the noise power of the receiver.
[0114] When the receiving point is connected to N devices, and N>1, the local computing speed can be derived as follows:
[0115]
[0116] in, For the wireless channel gain between different devices and receiving points, The effective capacitance coefficients for different devices.
[0117] When the complete uninstallation mode is selected, set Let be the time allocation factor for the unloading of the i-th device, then we have The uninstallation rate can be derived as follows:
[0118]
[0119] Assuming that within a time frame, only the wireless channel gain h... i It is time-varying, while other system parameters are fixed. Maximizing the weighted sum computation rate can be expressed as:
[0120]
[0121] Among them, w i As weight; This indicates the decision to unload.
[0122] Step four includes:
[0123] Store the current channel state h and the corresponding optimal offloading decision m as an experience (h, m) in the memory bank;
[0124] When the amount of experience stored in the memory bank reaches a certain number M, a batch of experience data is randomly sampled from the memory bank periodically for replay training.
[0125] To investigate the impact mechanism of different values of M on model performance, an ablation experiment was designed, and the results were compared as follows: Figure 2 As shown; Figure 2 (a) shows that the model performance fluctuates significantly when M=512. When the memory capacity is increased to M=1024, the model converges the fastest and has the best stability; further increasing the capacity to M=2048 can improve the final accuracy, but a significant increase in the number of training samples is required for convergence.
[0126] Based on the trade-off between convergence efficiency and stability, M was ultimately set to 1024.
[0127] The GCN network is trained using the empirical data described above. The network parameters are updated by minimizing the binary cross-entropy loss between the predicted unloading decision and the best storage decision, with the objective function value being minimized.
[0128] The present invention also provides an online computation offloading system for edge computing devices of graph neural networks, comprising:
[0129] Data preprocessing module: preprocesses the input data and constructs the node feature matrix of the edge computing device;
[0130] Graph Convolution Module: Constructs a GCN model, performs convolution processing on graph matrix feature data, and generates unloading probabilities;
[0131] Unloading decision generation module: Generates K candidate adaptive unloading decisions based on the unloading probability, evaluates the weighted calculation rate of each strategy, and selects the optimal unloading decision;
[0132] Replay Training Module: Based on the generated optimal unloading decision, the GCN model is updated through experience replay training.
Claims
1. An online computational offloading method for an edge computing device using a graph neural network, characterized in that, Includes the following steps: (1) Preprocess the input data to construct the node feature matrix of the edge computing device; (2) Construct a GCN model, perform convolution processing on the graph matrix feature data to generate unloading probabilities. The GCN model uses a two-layer GCN structure for feature extraction. The first layer maps the node features to 64 dimensions, and the second layer further maps the features to 32 dimensions, using ReLU as the activation function. in, For the input feature matrix, and These are the trainable weight matrices for the first and second layers, respectively; Finally, a fully connected layer is used in the output layer to map the 32-dimensional features to 1-dimensional features, and the sigmoid function is used to output the uninstallation probability for each user. in, Given the output layer weight matrix, the Sigmoid activation function formula is: ； (3) Generate based on unloading probability The system identifies several candidate adaptive unloading decisions, evaluates the weighted computation rate of each strategy, and selects the optimal unloading decision. (4) Based on the generated optimal unloading decision, the GCN model is updated by training through experience replay.
2. The online computational offloading method for edge computing devices using graph neural networks according to claim 1, characterized in that, In step (1), the input data is preprocessed to construct the node feature matrix of the edge computing device, including: Based on the number of edge computing devices Set their respective channel gain values. Based on Gaussian similarity, the association information between users is calculated, and an adjacency matrix is generated.
3. The online computational offloading method for edge computing devices using graph neural networks according to claim 2, characterized in that, The adjacency matrix formula is: in, The symmetric normalized adjacency matrix is calculated as follows: in, Adjacency matrix The degree matrix.
4. The online computational offloading method for edge computing devices using graph neural networks according to claim 1, characterized in that, In step (3), K candidate adaptive unloading decisions are generated based on the unloading probability, where K is an adjustable parameter with a maximum of N+1. The weighted computation rate of each strategy is evaluated, and the optimal unloading decision is selected, including: The first decision is obtained directly through binarization using a set threshold: values greater than the threshold are offloaded to the edge server, while values less than the threshold are computed locally. The remaining K-1 actions are to sequentially flip the decision of the least uncertain device based on the order in which the output value of each device is close to the threshold.
5. The online computational offloading method for edge computing devices using graph neural networks according to claim 4, characterized in that, The decision-making steps for the K candidate adaptive unloading are as follows: (1) Calculate the absolute distance between the unloading probability value of each device and the set threshold; (2) Sort the devices in ascending order based on their absolute distance, and put the least uncertain devices at the top to obtain the index list; (3) For the first K-1 most uncertain users, reverse the user's decision in turn to generate new candidate decisions; After obtaining K candidate decisions, calculate their corresponding weighted computation rate, and finally select the decision that brings the highest computation rate as the unloading decision for the current time frame through a binary evaluation method.
6. The online computational offloading method for edge computing devices using graph neural networks according to claim 1, characterized in that, In step (3), the weighted calculation rate is: The total system time is divided into several consecutive time frames of length T. In a static IoT environment, T is typically on the order of a few seconds, a relatively short time. It is assumed that the frame length is less than the channel coherence time. Each time frame can be divided into two phases: Phase 1: Downlink Wireless Power Transmission Phase: At the beginning of time frame T, the receiving point broadcasts radio frequency power for a duration of aT, where... This is the time allocation factor for energy transfer; the device harvests energy from the received radio frequency signal during this phase. Phase Two: Uplink Task Execution Phase: The remaining time (1-a)T is used for task execution and follows the binary offload strategy. The device must select one of the following two modes ( ): (a) Local computation (x=0): Computation is performed locally using the collected energy throughout the entire (1-a)T duration; (b) Complete offloading (x=1): The entire task is offloaded to the receiving point; the device uses the collected energy to upload the task data to the receiving point for remote execution; the time required for the receiving point to calculate the task and download the results back to the device is negligible compared to T. Let h represent the wireless channel gain between the receiving point and the device, and let the total energy collected by the device within one time frame be: in, For the efficiency of the energy harvesting circuit; Assume the device's processor has a computing speed of When local computing mode is selected, the energy consumed to execute computing tasks is Where k is the effective capacitance coefficient related to the chip structure, taken as... The local computing speed is: Assume the transmission power of the device is Then there is According to Shannon's formula, the uplink data transmission rate, i.e., the offload rate, is: in, For unified communication bandwidth, This refers to the noise power of the receiver; When the receiving point is connected to N devices, and N>1, the local computing rate is: in, For the wireless channel gain between different devices and receiving points, The effective capacitance coefficient for different devices; When the complete uninstallation mode is selected, set Let be the time allocation factor for the unloading of the i-th device, then we have The uninstallation rate can be derived as follows: Assuming that within a time frame, only the wireless channel gain h... i It is time-varying, while other system parameters are fixed. Maximizing the weighted sum computation rate can be expressed as: Among them, w i As weight; This indicates the decision to unload.
7. The online computational offloading method for edge computing devices using graph neural networks according to claim 1, characterized in that, In step (4), based on the generated optimal unloading decision, the GCN model is updated through experience replay training, including: The model maintains a fixed-size experience replay memory with a capacity of M; at each time frame, the model updates the current state decision pair. Stored in the buffer, where This refers to the channel gain between the corresponding device and the receiving point. The optimal decision is obtained through the binary search method; when the buffer is full, new experience will overwrite the oldest experience. The model employs a periodic training strategy. Every T time frames, a small batch of experience is randomly sampled from the experience replay buffer for training. Random sampling reduces the correlation between training samples and accelerates the convergence speed. The training process uses a binary cross-entropy loss function.
8. An online computational unloading system for an edge computing device using a graph neural network, characterized in that, include: Data preprocessing module: preprocesses the input data and constructs the node feature matrix of the edge computing device; Graph Convolution Module: Constructs a GCN model, performs convolution processing on graph matrix feature data, and generates unloading probabilities; Unloading decision generation module: Generates K candidate adaptive unloading decisions based on the unloading probability, evaluates the weighted calculation rate of each strategy, and selects the optimal unloading decision; Replay Training Module: Based on the generated optimal unloading decision, the GCN model is updated through experience replay training.
9. A computing device, characterized in that, include: One or more processors, one or more memories, and one or more programs, said programs being stored in the memory and configured to be executed by the processor, said programs, when loaded onto the processor, implementing the online computation offloading method for an edge computing device of a graph neural network according to any one of claims 1 to 7.
10. A storage medium, characterized in that, The storage medium stores a computer program, which includes program instructions that, when executed by a processor, cause the processor to perform the online computation offloading method for an edge computing device based on any one of claims 1 to 7 of the graph neural network.