Edge computing resource processing method and device, equipment, storage medium and product

Abstract

The invention discloses an edge computing resource processing method and device, equipment, a storage medium and a product, and relates to the technical field of edge computing, and the method comprises the steps: constructing a graph structure of edge computing nodes and user equipment, and initializing the features of the graph structure to obtain the node features of the graph structure; diffusing node features of the graph structure to obtain a diffused node feature matrix; inputting the diffused node feature matrix into a multi-layer perceptron to dynamically adjust node features, and generating new node feature representation; splicing the new node feature representation with a predetermined feature dimension identification matrix to generate an extended feature matrix; and carrying out resource allocation and scheduling according to the extended feature matrix. According to the method, better resource allocation and task scheduling can be realized, the utilization rate of edge computing resources is remarkably improved, the network delay is reduced, and the overall service quality is improved.

Description

Technical Field

[0001] The present application relates to the field of edge computing technology, and in particular to an edge computing resource processing method, device, equipment, storage medium and product. Background Art

[0002] With the development of 5G technology, edge computing has gradually become one of the key technologies to improve network performance and service quality. Edge computing deploys computing nodes close to users to achieve local processing and storage of data, thereby significantly reducing latency and improving computing efficiency. However, with the increasing number of edge computing nodes and user devices, how to efficiently allocate resources and schedule tasks in a complex network environment has become an urgent problem to be solved.

[0003] Existing technologies use GNN (Graph Neural Network) to perform edge computing resource allocation and task scheduling. GNN can utilize the connection relationship between nodes to perform efficient feature propagation and information fusion, which can help achieve resource optimization in complex network environments.

[0004] However, existing GNNs often fall into shallow local minima when dealing with large-scale and complex networks, resulting in inefficient resource scheduling. Summary of the invention

[0005] The main purpose of this application is to provide an edge computing resource processing method, device, equipment, storage medium and product, which can achieve better resource allocation and task scheduling in a complex network environment, and significantly improve the utilization rate of edge computing resources and resource scheduling efficiency.

[0006] To achieve the above objectives, the present application proposes an edge computing resource processing method, the method comprising:

[0007] Constructing a graph structure of edge computing nodes and user devices, and initializing features of the graph structure to obtain node features of the graph structure;

[0008] Diffusion of node features of the graph structure is performed to obtain a diffused node feature matrix;

[0009] Inputting the diffused node feature matrix into a multi-layer perceptron to dynamically adjust the node features and generate a new node feature representation;

[0010] The new node feature representation is concatenated with a predetermined feature dimension identification matrix to generate an extended feature matrix;

[0011] Resource allocation and scheduling are performed according to the extended feature matrix.

[0012] In one embodiment, the step of diffusing the node features of the graph structure to obtain a diffused node feature matrix includes:

[0013] The node features of the graph structure are diffused by using a heat kernel filter according to application scenario classification and combined with a Markov chain to obtain a diffused node feature matrix.

[0014] In one embodiment, the step of using a heat kernel filter to classify the node features of the graph structure according to the application scenario and combining the Markov chain to diffuse the node features of the graph structure to obtain a diffused node feature matrix includes:

[0015] Combined with different application scenarios of edge computing networks, heat kernel filters are classified to obtain the diffusion type of the heat kernel filters;

[0016] According to the diffusion type of the heat kernel filter, setting corresponding time parameters for the heat kernel filter;

[0017] Determine the corresponding heat kernel filter coefficient according to the time parameter of the heat kernel filter;

[0018] Determining a transfer probability matrix of the heat kernel filter based on the heat kernel filter coefficients and in combination with a Markov chain;

[0019] The node features of the graph structure are diffused according to the transition probability matrix to obtain a diffused node feature matrix.

[0020] In one embodiment, the step of determining the transfer probability matrix of the heat kernel filter based on the heat kernel filter coefficients and in combination with the Markov chain includes:

[0021] Determine a state vector of each node according to the node characteristics of the graph structure;

[0022] Construct the state space of the Markov chain based on the state vector of each node;

[0023] The transfer probability matrix of the heat kernel filter is calculated based on the heat kernel filter coefficients and in combination with the state space and network topology of the Markov chain.

[0024] In one embodiment, the step of calculating the transfer probability matrix of the heat kernel filter according to the heat kernel filter coefficients and in combination with the state space and network topology of the Markov chain includes:

[0025] Calculating the state influence weight of the neighboring nodes of the target node based on the state vectors of the target node and the neighboring nodes in the graph structure;

[0026] Determining the influence weight of the thermal kernel filter parameters according to the thermal kernel filter coefficients;

[0027] Determine the node state maintenance weight of the target node;

[0028] The transfer probability of the target node is calculated according to the adjacent node state influence weight, the heat kernel filter parameter influence weight and the node's own state retention weight;

[0029] Based on the state space of the Markov chain and the transition probability of each target node in the graph structure, a transition probability matrix of the heat kernel filter is calculated.

[0030] In one embodiment, the step of initializing the features of the graph structure to obtain the node features of the graph structure includes:

[0031] Initializing node features of the graph structure to obtain a node feature vector;

[0032] The initial node feature matrix is ​​obtained by constructing a matrix based on the node feature vectors of each node;

[0033] The step of diffusing the node features of the graph structure according to the transition probability matrix to obtain a diffused node feature matrix comprises:

[0034] The initial node feature matrix is ​​updated according to the transition probability matrix to obtain a diffused node feature matrix.

[0035] In one embodiment, the step of determining the feature dimension identification matrix includes:

[0036] Determining the characteristic dimension of the initial node characteristic matrix;

[0037] Assign an embedding vector to each feature dimension;

[0038] Based on the embedding vector and the embedding vector dimensions of the predetermined feature dimension identifiers, obtaining an initial feature dimension identifier matrix;

[0039] Training and optimizing the embedding vector in combination with the initial node feature matrix to obtain an optimized embedding vector;

[0040] An optimized feature dimension identification matrix is ​​determined based on the optimized embedding vector.

[0041] In one embodiment, the step of allocating resources according to the extended feature matrix includes:

[0042] Determine a task set, each task in the task set includes a computing requirement and a latency requirement;

[0043] Calculate the latency and resource usage cost of each task on each node;

[0044] Setting a resource allocation objective function according to a resource allocation goal, wherein the resource allocation goal is to minimize the delay and resource usage cost of the task and to meet the computing requirements and latency requirements of the task;

[0045] Determining an optimal resource allocation scheme according to the extended characteristic matrix to minimize the resource allocation objective function;

[0046] The corresponding tasks are allocated to the corresponding nodes according to the optimal resource allocation scheme.

[0047] In one embodiment, the step of performing resource scheduling according to the extended feature matrix includes:

[0048] Setting a task scheduling objective function according to the task scheduling goal, wherein the task scheduling goal is to minimize the total completion time of the task and maximize the utilization rate of computing resources;

[0049] Calculate the start time and execution time of each task;

[0050] According to the start time and execution time of each task, determine the execution order of each task to minimize the task scheduling objective function;

[0051] Resource scheduling is performed according to the determined execution order of each task.

[0052] The present application also proposes an edge computing resource processing device, the edge computing resource processing device comprising:

[0053] A construction module is used to construct a graph structure of edge computing nodes and user devices, and initialize the features of the graph structure to obtain node features of the graph structure;

[0054] A diffusion module, used for diffusing the node features of the graph structure to obtain a node feature matrix after diffusion;

[0055] An adjustment module, used for inputting the diffused node feature matrix into a multi-layer perceptron to dynamically adjust the node features and generate a new node feature representation;

[0056] A concatenation module, used for concatenating the new node feature representation with a predetermined feature dimension identification matrix to generate an extended feature matrix;

[0057] The resource allocation and scheduling module is used to perform resource allocation and scheduling according to the extended feature matrix.

[0058] The present application also proposes an edge computing resource processing device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the edge computing resource processing method as described above.

[0059] The present application also proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the edge computing resource processing method as described above are implemented.

[0060] The present application also proposes a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the edge computing resource processing method as described above.

[0061] One or more technical solutions proposed in this application have at least the following technical effects:

[0062] The edge computing resource processing method, device, equipment, storage medium and product proposed in this application, by constructing a graph structure of edge computing nodes and user devices, and initializing the features of the graph structure, obtain the node features of the graph structure; diffuse the node features of the graph structure to obtain a diffused node feature matrix; input the diffused node feature matrix into a multilayer perceptron to dynamically adjust the node features and generate a new node feature representation; splice the new node feature representation with a predetermined feature dimension identification matrix to generate an extended feature matrix; and perform resource allocation and scheduling according to the extended feature matrix. This application diffuses the node features of the graph structure, and inputs the diffused node feature matrix into a multilayer perceptron to dynamically adjust the node features, improve the accuracy and rationality of the feature representation, can achieve better resource allocation and task scheduling, significantly improve the utilization of edge computing resources, reduce network latency, improve the overall service quality, enhance the stability and adaptability of the system, support large-scale deployment, and the solution has a wide range of commercial application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0064] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0065] Figure 1 This is a flow chart of the first embodiment of the edge computing resource processing method of the present application;

[0066] Figure 2 This is a flow chart of the second embodiment of the edge computing resource processing method of the present application;

[0067] Figure 3 This is a schematic diagram of the overall process of the edge computing resource processing method of this application;

[0068] Figure 4 This is a schematic diagram of the module structure of the edge computing resource processing device according to an embodiment of the present application;

[0069] Figure 5 This is a schematic diagram of the device structure of the hardware operating environment involved in the edge computing resource processing method in the embodiment of the present application.

[0070] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0071] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0072] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0073] The main solution of the embodiment of the present application is: by constructing a graph structure of edge computing nodes and user devices, and initializing the features of the graph structure, the node features of the graph structure are obtained; the node features of the graph structure are diffused to obtain a diffused node feature matrix; the diffused node feature matrix is ​​input into a multilayer perceptron to dynamically adjust the node features and generate a new node feature representation; the new node feature representation is spliced ​​with a predetermined feature dimension identification matrix to generate an extended feature matrix; and resources are allocated and scheduled according to the extended feature matrix. Among them, the node features of the graph structure are diffused by using a hot kernel filter according to the application scenario classification and combined with a Markov chain to obtain a diffused node feature matrix. The present application diffuses the node features of the graph structure, and inputs the diffused node feature matrix into a multilayer perceptron to dynamically adjust the node features, improve the accuracy and rationality of the feature representation, and can achieve better resource allocation and task scheduling, significantly improve the utilization of edge computing resources, reduce network latency, improve the overall service quality, enhance the stability and adaptability of the system, and support large-scale deployment. The solution has a wide range of commercial application value.

[0074] This embodiment takes into account that traditional resource allocation and task scheduling methods, such as centralized management, rule-based optimization, etc., are often difficult to cope with large-scale and highly dynamic network environments. For example, centralized management, although it has certain management convenience, exposes many disadvantages in large-scale and highly dynamic network environments. Centralized management relies on a single central control node, which needs to process massive amounts of information. As the scale of the network expands, the pressure of information processing increases dramatically, and it is very easy to become a system bottleneck. At the same time, it lacks in-depth mining of the network topology, and cannot flexibly allocate resources based on the complex connection relationship between nodes, making it difficult to achieve global optimal configuration. The rule-based optimization method is difficult to adapt to the dynamic changes of the network due to the relative fixedness of the rules. When faced with frequent node additions and subtractions or load fluctuations, it is impossible to adjust the strategy in time, resulting in a long-term low level of computing resource utilization, and network latency also increases, which seriously affects the overall network performance and user experience. These methods usually lack full utilization of the network topology, and cannot achieve the optimal configuration of resources on a global scale, resulting in low computing resource utilization, high network latency, and other problems.

[0075] To solve this problem, graph neural networks (GNNs) are gradually being used in edge computing resource allocation and task scheduling due to their unique advantages in processing graph structured data. GNNs can use the connection relationship between nodes to perform efficient feature propagation and information fusion, helping to achieve resource optimization in complex network environments. However, when dealing with large-scale and complex networks, existing GNNs are often prone to falling into shallow local minima, resulting in poor optimization results.

[0076] Therefore, this application provides a solution and proposes a 5G edge computing resource allocation and scheduling method, which combines the stability of the thermal kernel filter and the nonlinear expression ability of the multilayer perceptron, effectively attracts the optimization trajectory to reach the deep local minimum in a complex network environment, and achieves better resource allocation and task scheduling, thereby significantly improving the utilization of edge computing resources, reducing network latency, and improving the overall service quality.

[0077] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device that can realize the above functions.

[0078] Based on this, the embodiment of the present application provides an edge computing resource processing method, referring to Figure 1 , Figure 1 This is a flow chart of the first embodiment of the edge computing resource processing method of the present application.

[0079] like Figure 1 As shown, in this embodiment, the edge computing resource processing method includes steps S10 to S50.

[0080] Each step is explained in detail below.

[0081] Step S10, constructing a graph structure of edge computing nodes and user devices, and initializing features of the graph structure to obtain node features of the graph structure;

[0082] This embodiment can be applied to a 5G network or other networks, and the edge computing node can be an edge computing node of a 5G network or other networks.

[0083] The graph structure includes a network topology structure composed of user devices and edge computing nodes.

[0084] The graph structure includes nodes and edges, and correspondingly, the features of the graph structure include node features and edge features. Each node in the graph structure constitutes a node set, and each edge in the graph structure constitutes an edge set.

[0085] Among them, the node set includes edge computing nodes and user devices, wherein the edge computing nodes refer to computing nodes distributed at the edge of the network, which are used to process data and computing tasks of user devices; the user devices refer to terminal devices connected to the edge computing nodes, which are used to send data and computing requests to the edge computing nodes.

[0086] The edge set consists of communication links between nodes, such as the communication link between an edge computing node and a user device, the communication link between two edge computing nodes, etc.

[0087] In this embodiment, the graph structure constructed with the above nodes and edges is initialized, and the node features of the graph structure are obtained by initializing the features of the graph structure.

[0088] Among them, a graph structure of edge computing nodes and user devices is constructed, and the features of the graph structure are initialized to obtain the node features of the graph structure, which can provide a basic data framework for subsequent operations.

[0089] As an implementation manner, the step of initializing the features of the graph structure and obtaining the node features of the graph structure may include:

[0090] Initialize the node features of the graph structure to obtain a node feature vector; construct a matrix based on the node feature vector of each node to obtain an initial node feature matrix. Subsequently, the node features of the graph structure can be diffused based on the initial node feature matrix.

[0091] Specifically, at the time of initialization, the feature vector of each node represents the properties and status of the node. The node features include: computing power, such as CPU, GPU, etc.; storage capacity, such as memory, hard disk, etc.; current load; location, that is, the geographical location of the node. Then, the matrix composed of the feature vectors of all nodes is set as the initial node feature matrix.

[0092] Step S20, diffusing the node features of the graph structure to obtain a node feature matrix after diffusion;

[0093] As an implementation method, a heat kernel filter can be used to classify the node features of the graph structure according to the application scenario and combined with a Markov chain to diffuse the node features to obtain a diffused node feature matrix. Among them, using a heat kernel filter to classify the application scenario and combine it with a Markov chain to diffuse the features can stably and accurately propagate features and optimize strategies.

[0094] Specifically, after constructing and initializing the graph structure, this scheme uses a heat kernel filter to diffuse the node features of the graph nodes to ensure stable propagation and homogenization of the features in the graph structure.

[0095] Furthermore, considering the complex and changeable environment of the 5G edge computing network, in order to make the feature diffusion process more accurate and in line with the actual network conditions, this solution introduces a Markov chain in the diffusion process of node features. Different probability transfer matrices can be set according to the characteristics of different types of heat kernel filters, and the node features of the graph structure can be diffused according to the probability transfer matrix to obtain the diffused node feature matrix. In this way, the functions and related parameters of the heat kernel filter can be fully utilized, so that the propagation changes of node features in the network can better reflect the network dynamics, thereby optimizing resource allocation and task scheduling strategies.

[0096] Step S30, inputting the diffused node feature matrix into a multilayer perceptron to dynamically adjust the node features and generate a new node feature representation;

[0097] Among them, dynamic adjustment of node features through multi-layer perceptron can improve model capabilities and approach the optimal solution.

[0098] As an implementation method, the diffused node feature matrix can be input into a multilayer perceptron to generate a new node feature representation through nonlinear transformation and feature combination.

[0099] The multilayer perceptron in this solution is a feedforward neural network consisting of multiple hidden layers that can achieve complex nonlinear mapping. MLP improves the model's expressiveness and generalization capabilities by learning nonlinear combinations of features layer by layer.

[0100] Step S40, concatenating the new node feature representation with a predetermined feature dimension identification matrix to generate an extended feature matrix;

[0101] After the diffused node feature matrix is ​​input into the multilayer perceptron, a new node feature representation is generated through nonlinear transformation and feature combination to obtain a new node feature representation. The new node feature representation is a multidimensional node feature. In this embodiment, the multidimensional node feature is flattened into a one-dimensional vector and concatenated with the feature dimension identifier embedding to form an extended feature matrix.

[0102] In this step, the new multi-dimensional node features are flattened into one-dimensional vectors and concatenated with the feature dimension identifier embedding to simplify the feature processing process and improve computational efficiency while preserving the relationship between features to ensure the effectiveness of the optimization process.

[0103] In specific implementation, the input features are converted into low-dimensional representations through the linear transformation layer, and the output features are converted back to the output dimension. After the input features are processed by the linear transformation layer with dimensionality reduction, the intermediate result of the low-dimensional representation is obtained. However, subsequent steps (such as feature concatenation and other operations) expect the features to have a specific dimensional format for effective calculation and processing. Therefore, it is necessary to convert the low-dimensional output features obtained through linear transformation back to the output dimension suitable for subsequent steps to ensure the coherence and compatibility of the data in the entire process. The role of this step is to achieve efficient feature mapping through linear transformation and ensure the consistency of the input and output feature dimensions.

[0104] Among them, the multi-dimensional node features are flattened into one-dimensional vectors to facilitate subsequent linear transformation and processing.

[0105] In this embodiment, feature dimension identification embedding is introduced, and the feature dimension identification is spliced ​​with the flattened feature vector to form an extended feature matrix. The feature dimension identification matrix is ​​introduced in this step to retain the relationship and semantic information between features, so as to enhance the effect of feature processing and improve the accuracy of feature mapping. Among them, the feature dimension identification matrix can retain the relationship and semantic information between features, enhance the effect of feature processing, and accurately represent the semantic information of feature dimensions, improve the accuracy of feature mapping. In addition, the feature dimension identification matrix can also be dynamically adjusted according to task requirements and network environment, and adapt to different optimization goals, thereby ensuring the efficiency and stability of feature processing and optimization process. By optimizing the feature dimension identification matrix, the accuracy of feature mapping can be significantly improved, providing a solid foundation for subsequent resource allocation and task scheduling.

[0106] Step S50: performing resource allocation and scheduling according to the extended feature matrix.

[0107] Among them, by expanding the feature matrix, the best resource allocation and task scheduling strategy can be formulated, thereby achieving better resource allocation and task scheduling.

[0108] Among them, when formulating the best resource allocation strategy, the task set and resource allocation goals are considered. Each task in the task set contains computing requirements and latency requirements, and the resource allocation goal is to minimize the task delay and resource usage cost and meet the computing requirements and latency requirements of the task.

[0109] In addition, in order to achieve effective resource allocation, this embodiment also introduces a resource allocation objective function. After considering task delay and resource usage cost, the resource allocation objective function is set.

[0110] The resource allocation process includes:

[0111] First, the delay and resource usage cost of each task on each node are calculated, all possible allocation schemes are compared, and the allocation scheme that minimizes the resource allocation objective function is selected. Then the task is assigned to the selected node.

[0112] Among them, when determining the task scheduling strategy, it is necessary to determine the execution order of tasks to ensure that tasks can be completed on time and maximize resource utilization.

[0113] When developing the best resource scheduling strategy, the task set and task scheduling goals are considered. The task scheduling goals are to minimize the total completion time of the task and maximize the utilization of computing resources. The resource scheduling process may include:

[0114] First, for each task, calculate its possible start time and execution time, then compare all possible scheduling sequences, select the scheduling sequence that minimizes the task scheduling objective function, then determine the start time and end time of the task, and update the task scheduling table.

[0115] Specifically, as an implementation mode, the step of allocating resources according to the extended feature matrix may include:

[0116] Determine a task set, each task in the task set includes a computing requirement and a latency requirement;

[0117] Calculate the latency and resource usage cost of each task on each node;

[0118] Setting a resource allocation objective function according to a resource allocation goal, wherein the resource allocation goal is to minimize the delay and resource usage cost of the task and to meet the computing requirements and latency requirements of the task;

[0119] Determining an optimal resource allocation scheme according to the extended characteristic matrix to minimize the resource allocation objective function;

[0120] The corresponding tasks are allocated to the corresponding nodes according to the optimal resource allocation scheme.

[0121] Specifically, as an implementation mode, the step of performing resource scheduling according to the extended feature matrix may include:

[0122] Setting a task scheduling objective function according to the task scheduling goal, wherein the task scheduling goal is to minimize the total completion time of the task and maximize the utilization rate of computing resources;

[0123] Calculate the start time and execution time of each task;

[0124] According to the start time and execution time of each task, determine the execution order of each task to minimize the task scheduling objective function;

[0125] Resource scheduling is performed according to the determined execution order of each task.

[0126] In practical applications, network environments and task requirements change dynamically. Therefore, dynamic adjustment and real-time optimization are required to ensure the efficiency of resource allocation and task scheduling. For example, according to real-time monitoring data, resource allocation and task scheduling strategies are dynamically adjusted. When changes in network status or task requirements are detected, resource allocation and task scheduling schemes are recalculated.

[0127] In this embodiment scheme, we can first use graph structure modeling to make full use of network topology information to provide a basis for subsequent feature diffusion and optimization, and use hot kernel filters to diffuse and smooth node features to generate a diffused feature matrix to ensure that negative feedback is not introduced during feature diffusion and avoid instability during the optimization process. Then, through the nonlinear mapping capability of the multilayer perceptron, the feature representation is dynamically adjusted to gradually approach the optimal solution, reducing task delays and resource usage costs, and the obtained multidimensional features are flattened into a one-dimensional vector, and a feature dimension identification matrix is ​​introduced to retain the relationship between features and semantic information, which can not only improve the efficiency of feature processing, but also improve the effect of feature processing. Finally, according to the optimized feature representation, computing resources are reasonably allocated, and task scheduling strategies are set.

[0128] Through the organic combination of the above steps, this solution can effectively attract the optimization trajectory to reach the deep local minimum in a complex network environment, thereby achieving better resource allocation and task scheduling. The careful design and mutual coordination of each step ensure the efficiency and stability of the optimization process, providing strong technical support for 5G edge computing.

[0129] Based on the first embodiment of the present application, the second embodiment of the present application is proposed. In the second embodiment of the present application, the same or similar contents as those of the above-mentioned first embodiment can be referred to the above introduction, and will not be repeated in the following.

[0130] On this basis, the step of using a heat kernel filter to classify the application scenarios and combining the Markov chain to diffuse the node features of the graph structure to obtain a diffused node feature matrix includes steps A10 to A50. Among them:

[0131] A10, classifying the heat kernel filter in combination with different application scenarios of the edge computing network to obtain the diffusion type of the heat kernel filter;

[0132] Among them, in the classification of heat kernel filters, combined with different application scenarios of edge computing networks, heat kernel filters are divided into fast diffusion type, medium diffusion type and slow diffusion type. Among them, the fast diffusion type heat kernel filter can make node features propagate over a larger range in a shorter time, which is suitable for scenarios where network topology changes frequently and requires fast response; the slow diffusion type pays more attention to the local stability and accuracy of features, and is suitable for situations where the network state is relatively stable and has high requirements for feature accuracy; the medium diffusion type is between the two and can be used as a balanced choice.

[0133] A20, according to the diffusion type of the heat kernel filter, setting corresponding time parameters for the heat kernel filter;

[0134] According to different diffusion types of the heat kernel filter, different time parameter values ​​can be set for the heat kernel filter to determine the corresponding heat kernel filter coefficient values.

[0135] A30, determining a corresponding heat kernel filter coefficient according to a time parameter of the heat kernel filter;

[0136] A40, determining a transfer probability matrix of the heat kernel filter based on the heat kernel filter coefficients and in combination with a Markov chain;

[0137] Specifically, as an implementation method, the state vector of each node can be determined according to the node characteristics of the graph structure; then the state space of the Markov chain is constructed according to the state vector of each node; and then the transfer probability matrix of the hot kernel filter is calculated according to the hot kernel filter coefficient and combined with the state space of the Markov chain and the network topology structure.

[0138] Wherein, as an implementation mode, the step of calculating the transfer probability matrix of the heat kernel filter according to the heat kernel filter coefficient and in combination with the state space and network topology structure of the Markov chain may include:

[0139] First, based on the state vectors of the target node and the adjacent nodes in the graph structure, the adjacent node state influence weight of the target node is calculated;

[0140] Then, determining the influence weight of the heat kernel filter parameters according to the heat kernel filter coefficient;

[0141] Then, determining the node state maintenance weight of the target node;

[0142] Then, the transfer probability of the target node is calculated according to the adjacent node state influence weight, the heat kernel filter parameter influence weight and the node's own state retention weight;

[0143] Finally, based on the state space of the Markov chain and the transition probability of each target node in the graph structure, the transition probability matrix of the heat kernel filter is calculated.

[0144] A50, diffusing the node features of the graph structure according to the transition probability matrix to obtain a diffused node feature matrix.

[0145] Specifically, as an implementation method, after obtaining a node feature vector by initializing the node features of the graph structure, and constructing a matrix based on the node feature vectors of each node to obtain an initial node feature matrix, the initial node feature matrix can be updated according to the transition probability matrix to obtain a diffused node feature matrix.

[0146] In this embodiment, after building and initializing the graph structure, this solution uses a hot kernel filter to diffuse the node features of the graph nodes to ensure the stable propagation and homogenization of the features in the graph structure. Considering the complex and changeable environment of the 5G edge computing network, in order to make the feature diffusion process more accurate and in line with the actual network conditions, this solution introduces a Markov chain in the diffusion process of node features, and sets different probability transfer matrices according to the characteristics of different types of hot kernel filters. In this way, the functions and related parameters of the hot kernel filter are fully utilized, so that the propagation changes of node features in the network can better reflect the network dynamics, thereby optimizing resource allocation and task scheduling strategies.

[0147] Based on the first embodiment or the second embodiment of the present application, the third embodiment of the present application is proposed. In the third embodiment of the present application, the same or similar contents as those of the first embodiment or the second embodiment can be referred to the above introduction, and will not be repeated in the following.

[0148] On this basis, please refer to Figure 2 , Figure 2 A flowchart diagram is provided for the third embodiment of the edge computing resource processing method of the present application.

[0149] like Figure 2As shown, in this embodiment, the step of determining the feature dimension identification matrix includes:

[0150] Step S401, determining the feature dimension of the initial node feature matrix;

[0151] Step S402, assigning an embedding vector to each feature dimension;

[0152] Step S403, obtaining an initial feature dimension identification matrix based on the embedding vector and the embedding vector dimension of the predetermined feature dimension identification;

[0153] Step S404, training and optimizing the embedding vector in combination with the initial node feature matrix to obtain an optimized embedding vector;

[0154] Step S405: determining an optimized feature dimension identification matrix based on the optimized embedding vector.

[0155] In this embodiment, feature dimension identifier embedding is introduced, and the feature dimension identifier is concatenated with the flattened feature vector to form an extended feature matrix. The feature dimension identifier matrix is ​​introduced in this step to retain the relationship and semantic information between features, so as to enhance the effect of feature processing and improve the accuracy of feature mapping.

[0156] The feature dimension identification matrix is ​​an embedding matrix used to represent the semantic information of each feature dimension. For each feature dimension, an embedding vector is assigned to it. These embedding vectors can capture the relationship between features and provide useful information in the feature processing process.

[0157] The feature dimension identification matrix can be determined by the following steps:

[0158] First, determine the feature dimension of the initial node feature matrix and determine the embedding vector dimension identified by the feature dimension;

[0159] Then, an embedding vector is assigned to each feature dimension;

[0160] Then, based on the embedding vector and the embedding vector dimension of the predetermined feature dimension identifier, an initial feature dimension identifier matrix is ​​obtained;

[0161] Then, the embedding vector is trained and optimized in combination with the initial node feature matrix to obtain an optimized embedding vector;

[0162] The embedding vector can be optimized through the model training process to better represent the semantic information of the feature dimension. In the model training process, the loss function is combined for iterative optimization. In each iteration, the gradient of the loss function is calculated through the back propagation algorithm and the embedding vector is updated. The embedding vector is updated according to the feedback of the loss function and gradually adjusted to the optimal state.

[0163] Finally, the optimized feature dimension identification matrix is ​​determined based on the optimized embedding vector.

[0164] During the model training process, the feature dimension identification matrix will be optimized based on the feedback of the loss function.

[0165] Among them, the feature dimension identification matrix can retain the relationship and semantic information between features, enhance the effect of feature processing, accurately represent the semantic information of feature dimensions, and improve the accuracy of feature mapping. In addition, the feature dimension identification matrix can also be dynamically adjusted according to task requirements and network environment to adapt to different optimization goals, thereby ensuring the efficiency and stability of feature processing and optimization process. By optimizing the feature dimension identification matrix, the accuracy of feature mapping can be significantly improved, providing a solid foundation for subsequent resource allocation and task scheduling.

[0166] The following takes the resource allocation and scheduling method of 5G network edge computing resources as an example, combined with Figure 3 , the implementation method of this application is elaborated in detail:

[0167] In the present application scheme, firstly, a graph structure of edge computing nodes and user devices is constructed and features are initialized to provide a basic data framework for subsequent operations; then, a hot kernel filter is used to classify the application scenarios and a Markov chain is used to diffuse the features in order to stably and accurately propagate the features and optimize the strategies; then, the node features are dynamically adjusted through a multi-layer perceptron to improve the model capabilities and approach the optimal solution; the multi-dimensional node features are flattened and embedded with the feature dimension identifiers to simplify the processing process and retain the feature relationships; finally, a resource allocation and task scheduling strategy is formulated based on the optimized features, while taking into account factors such as task delay and cost, to achieve better resource allocation and task scheduling in a complex network environment, effectively improving the utilization of edge computing resources, reducing network delays and improving the overall service quality.

[0168] like Figure 3 As shown, Figure 3 This is a schematic diagram of the main process of this application plan, which includes five steps.

[0169] The following are detailed descriptions of these five steps:

[0170] Step 1: Model the graph structure for edge computing nodes and user devices in the 5G network and perform feature initialization.

[0171] In this scheme, the edge computing nodes and user devices in the 5G network are first modeled as a graph structure G = (V, E), and the features of the nodes and edges are initialized.

[0172] Specifically, the node set V includes edge computing nodes v i and user device u j , where the edge computing node v i Represents computing nodes distributed at the edge of the network, used to process data and computing tasks of user devices; user devices u j Represents a terminal device connected to an edge computing node, used to send data and computing requests to the edge computing node.

[0173] The edge set E consists of communication links between nodes, such as the communication link between an edge computing node and a user device, the communication link between two edge computing nodes, etc. ij To represent, it means node v i and node u j The links between nodes, or i and node v j The link between.

[0174] The adjacency matrix M can be set as a |V|×|V| matrix, where M ij Represents node v i and node v j The connection relationship between them, if the node v i and node v j If there is a direct communication link between ij =1, otherwise M ij =0.

[0175] Then, the graph structure constructed above is initialized. During initialization, each node v i The eigenvector h i Represents the properties and status of the node. Node characteristics include the following: computing capacity CC, such as CPU, GPU, etc.; storage capacity SC, such as memory, hard disk, etc.; current load CL; location L, that is, the geographical location of the node.

[0176] Thus, the node feature vector h i It can be expressed as:

[0177] h i =[CC i ,SC i ,CL i ,L i ]

[0178] The matrix composed of the feature vectors of all nodes is set as the initial node feature matrix H. The initial node feature matrix H is a matrix of |V|×d, where |V| is the number of nodes and d is the feature dimension.

[0179] Then the initial node feature matrix H is:

[0180]

[0181] In the graph structure constructed above, each edge e ij The eigenvector m ij Indicates the properties and status of the communication link. The edge features include the following: bandwidth BW; delay LAT; link quality LQ.

[0182] Thus, the edge eigenvector m ij It can be expressed as:

[0183] m ij =[BW ij ,LAT ij ,LQ ij ]

[0184] The matrix composed of the eigenvectors of all edges is set as the edge feature matrix M, which is a matrix of |E|×d, where |E| is the number of edges and d is the feature dimension.

[0185] The degree matrix D is a diagonal matrix where D ii Represents node v i The degree of v i The number of directly connected edges. ii It can be expressed as:

[0186]

[0187] in, Represents the symmetric normalized adjacency matrix. Through normalization, the feature diffusion process can be made more stable and the imbalance caused by the difference in node degrees can be avoided.

[0188] Step 2: Use the heat kernel filter to diffuse the node features of the graph nodes.

[0189] After building and initializing the graph structure, this solution uses a hot kernel filter to diffuse the node features of the graph nodes to ensure the stable propagation and homogenization of the features in the graph structure. Considering the complex and changeable environment of the 5G edge computing network, in order to make the feature diffusion process more accurate and in line with the actual network conditions, this solution introduces a Markov chain in the diffusion process of node features, and sets different probability transfer matrices according to the characteristics of different types of hot kernel filters. In this way, the functions and related parameters of the hot kernel filter are fully utilized, so that the propagation changes of node features in the network can better reflect the network dynamics, thereby optimizing resource allocation and task scheduling strategies.

[0190] This step includes three parts: (1) classification and setting of heat kernel filters; (2) determination of the transfer probability matrix based on different heat kernel filters; (3) feature diffusion based on Markov chain.

[0191] (1) Classification and setting of thermal kernel filters:

[0192] In the classification and setting of heat kernel filters, combined with different application scenarios of edge computing networks, heat kernel filters are divided into fast diffusion type, medium diffusion type and slow diffusion type. Among them, the fast diffusion type heat kernel filter can propagate node features in a larger range in a shorter time, which is suitable for scenarios where the network topology changes frequently and requires fast response; the slow diffusion type pays more attention to the local stability and accuracy of the features, and is suitable for situations where the network state is relatively stable and the feature accuracy requirements are high; the medium diffusion type is between the two and can be used as a balance option. For different diffusion types of heat kernel filters, different time parameter values ​​can be set for the heat kernel filter to determine the corresponding heat kernel filter coefficient value.

[0193] Among them, the heat kernel filter coefficient is f n , which is used to control the contribution of adjacency matrix powers with different hop numbers n to feature diffusion, that is, n is the order of the adjacency matrix power.

[0194]

[0195] Here, t is the diffusion time parameter.

[0196] In some scenarios with extremely high real-time requirements, such as the emergency brake signal transmission scenario in the Internet of Vehicles, information needs to be widely disseminated in a short time. For example, data transmission and feature update are required every 10 milliseconds. Therefore, a fast diffusion type heat kernel filter can be set. The time parameter t of the heat kernel filter can be set to 0.05. At this time, the heat kernel filter coefficient

[0197] In this fast-update scenario, for lower-order adjacency matrix powers, the heat kernel filter coefficients are relatively large, which enables node features to propagate rapidly between adjacent nodes in a very short time, quickly affecting the states of surrounding nodes to meet the needs of rapid diffusion of emergency braking signals.

[0198] In some scenarios where time is not a big requirement, such as in the production monitoring system of a smart factory, data needs to be transmitted in a timely manner within a certain range, but it also needs to maintain a certain stability, such as data transmission and feature update every 100 milliseconds. Therefore, a medium-speed diffusion type heat kernel filter can be used. The time parameter t of the heat kernel filter can be set to 0.3. At this time, the heat kernel filter coefficient

[0199] In this scenario, the distribution of heat kernel filter coefficients on each order of adjacency matrix power is relatively balanced, which can not only achieve a faster diffusion of node features in a certain area within a time interval of 100 milliseconds, but also avoid instability due to excessive diffusion, ensuring timely and stable dissemination of production equipment status information.

[0200] In some scenarios that require extremely high data accuracy and stability, such as financial transaction data processing networks, features need to diffuse slowly and accurately, such as data transmission and feature update every 500 milliseconds. Therefore, a slow diffusion type heat kernel filter is used. The time parameter t of the heat kernel filter can be set to 1.2. At this time, the heat kernel filter coefficient

[0201] In this scenario, the coefficients of the heat kernel filters of higher-order adjacency matrix powers are relatively more significant, and node features are mainly propagated gradually through multi-hop neighbor nodes, focusing on the deep propagation of features and the long-term stability of the overall network to ensure the accuracy and stability of financial transaction data processing.

[0202] (2) Determination of the transfer probability matrix based on different heat kernel filters:

[0203] In the determination of the transfer probability matrix based on different heat kernel filters, the state space of the Markov chain can be defined first. Based on the node attributes and states, computing power (CC), storage capacity (SC), current load (CL) and location (L) are selected as the key features for constructing the state space of the Markov chain, and each key feature is discretized, such as dividing computing power into three levels: high, medium and low; dividing storage capacity into different intervals according to the size of memory and hard disk capacity; dividing the current load into several levels in percentage form, and classifying the location according to the relative position in the geographical area or network topology.

[0204] Then, the state of each node is represented by a state vector, such as for node v i , whose state vector in, They are the discretized representations of computing power, storage capacity, current load, and location, respectively.

[0205] The state space Ω of the entire network consists of all possible combinations of node state vectors. The size of the state space depends on the number of discretization levels of each feature. For example, if each feature has n discrete levels, then for the above four features, the state space size is n 4 .

[0206] After defining the state space, we can determine the transition probability matrix. Let It represents the probability of transitioning from state i to state j under the kth heat kernel filter (k=1 is fast diffusion type, k=2 is medium diffusion type, k=3 is slow diffusion type). The transition probability matrix P k The calculation is based on factors such as node status characteristics, heat kernel filter characteristics, and network topology. The general formula is:

[0207]

[0208] in is a weight factor related to state transition, which comprehensively considers the influence of factors such as adjacent node states, heat kernel filter parameters and node states on the transition probability, and |Ω| is the size of the state space. The following determines the corresponding escape probability for heat kernel filters of different diffusion types:

[0209] a) Calculation of transfer probability of fast diffusion heat kernel filter

[0210] 1. Adjacent node status affects weight

[0211] For adjacent nodes v m With the target node v n ,set up and are their state vectors respectively. Define the influence weight of adjacent node states for:

[0212]

[0213] Among them, β f is the weight coefficient of feature f (f = 1 for computing power, f = 2 for storage capacity, f = 3 for current load, f = 4 for location), which is pre-set according to the importance of the feature to the state transition (e.g. β1 = 0.3, β2 = 0.2, β3 = 0.4, β4 = 0.1). max f and min fare the maximum and minimum values ​​of feature f in the state space. For example, the value range of the computing power feature in the state space is [0,10] (assuming it is represented by some normalized numerical value), then max1 = 10, min1 = 0. When the state difference between adjacent nodes is large, The larger the value, the higher the probability that the target node will transfer to the state of the adjacent node.

[0214] 2. Thermal Kernel Filter Parameters Influence Weights

[0215] According to the fast diffusion heat kernel filter coefficient It is possible to define the influence weights of the heat kernel filter parameters for:

[0216]

[0217] For lower order adjacency matrix powers (such as n = 1, 2), The value is relatively large, emphasizing the rapid feature propagation between adjacent nodes. For example, when n = 1, When n=2,

[0218] 3. The node’s own state maintains weight

[0219] Define the node's own state to maintain weight for:

[0220]

[0221] in, is the target node v n The discretized value of the current load. For example, the current load is in the range of [0,5] after discretization. but When the node load is low, The larger the value, the more likely the node is to maintain the current state, otherwise it is more likely to transfer state.

[0222] 4. Calculate the transfer probability by comprehensive weight

[0223] Combining the above weights, the weight factor The calculation is as follows:

[0224]

[0225] Among them, N n is node v n The set of adjacent nodes, N is the maximum order of the adjacency matrix power considered (e.g., N = 2). Finally, the transition probability can be calculated according to the above general formula

[0226]

[0227] b) Calculation of transfer probability of medium-speed diffusion heat kernel filter

[0228] 1. Adjacent node status affects weight

[0229] Adjacent node status affects weight The calculation method is similar to the fast diffusion type, but the characteristic weight coefficient is adjusted to β f (For example, β1 = 0.25, β2 = 0.25, β3 = 0.3, β4 = 0.2) to meet the needs of medium-speed diffusion:

[0230]

[0231] 2. Thermal Kernel Filter Parameters Influence Weights

[0232] According to the coefficient of the medium-speed diffusion heat kernel filter It is possible to define the influence weights of the heat kernel filter parameters for:

[0233]

[0234] The weight distribution of each order of adjacency matrix power is relatively balanced, achieving medium-speed and stable feature diffusion. For example, when n = 1, When n=2, When n=3,

[0235] 3. The node’s own state maintains weight

[0236] The node's own state maintains the weight for:

[0237]

[0238] The impact of node load on state retention is also considered, but the impact is lower than that of the rapid diffusion type. but

[0239] 4. Calculate the transfer probability by comprehensive weight

[0240] Weight Factor The calculation is as follows:

[0241]

[0242] Finally, the transition probability is calculated according to the general formula above

[0243] c) Calculation of transfer probability of slow diffusion heat kernel filter

[0244] 1. Adjacent node status affects weight

[0245] Adjacent node status affects weight The calculation method is similar to the fast diffusion type, but the characteristic weight coefficient is adjusted to β again f (e.g. β1 = 0.2, β2 = 0.2, β3 = 0.35, β4 = 0.25), and pay more attention to the long-term impact of characteristics such as position and load on state transition:

[0246]

[0247] 2. Thermal Kernel Filter Parameters Influence Weights

[0248] According to the slow diffusion heat kernel filter coefficient It is possible to define the influence weights of the heat kernel filter parameters for:

[0249]

[0250] For higher order adjacency matrix powers (such as n=3,4), A larger value emphasizes the deep propagation of features. For example, when n = 3, When n=4,

[0251] 3. The node’s own state maintains weight

[0252] The node's own state maintains the weight for:

[0253]

[0254] The node load has a relatively small impact on state retention, and more emphasis is placed on long-term stable state trends. For example, if but

[0255] 4. Comprehensive weight calculation and transfer probability determination

[0256] Weight Factor The calculation is as follows:

[0257]

[0258] Finally, the transition probability is calculated according to the general formula above

[0259] Afterwards, based on the transition probability of each node in the graph structure, the transition probability matrix of the heat kernel filter is calculated.

[0260] (3) Feature diffusion based on Markov chain:

[0261] Based on the initial node feature matrix H in step 1, according to the initial type and parameters of the heat kernel filter (such as the time parameter t value), the initial feature vector of each node is converted into the corresponding initial state vector to determine the initial state S0 of the network.

[0262] For the fast diffusion heat kernel filter (t = 0.05), when setting the initial state, according to the heat kernel filter coefficient (When n=0, f0=e -0.05 ≈0.9512), make preliminary adjustments to the node features. For example, for node v i Computational power characteristics of CC i , the initial state is adjusted to Other features are similarly adjusted to reflect the trend of feature changes in the early stage of rapid diffusion, that is, the node features have been affected by diffusion to a certain extent in the initial state.

[0263] For the medium diffusion heat kernel filter (t = 0.3) and the slow diffusion heat kernel filter (t = 1.2), the corresponding heat kernel filter coefficients f n Similar adjustments are made to the initial node features.

[0264] After setting the initial state, the state transfer and feature update are performed in combination with the heat kernel filter type and parameters, including:

[0265] 1. At each time step t, according to the current network state S t And the transition probability matrix P corresponding to the heat kernel filter type k (k=1, 2, 3 correspond to fast, medium and slow diffusion types respectively), calculate the network state S of the next time step t+1 , the calculation formula is:

[0266]

[0267] in is the probability of being in state i at the current time step t (i.e., the state vector S t ), is the probability of transitioning from state i to state j under the kth heat kernel filter.

[0268] 2. According to the new network status S t+1 The initial node feature matrix H is updated by taking into account the continuous effect of the heat kernel filter (such as the continuous influence of the change of the time parameter t on the feature diffusion).

[0269] For the fast diffusion heat kernel filter, as the time step t increases, the heat kernel filter coefficient Constantly changing. When updating node features, for example, for node v i Computational power characteristics of CC i , the update formula is:

[0270]

[0271] in, It reflects the influence of the change of heat kernel filter coefficient over time on node characteristics. Other characteristics (such as storage capacity SC i , Current load CL i and position L i ) also adopts a similar updating method and adjusts according to their respective feature weights and state transition probabilities to reflect the dynamic changes of node features under the fast diffusion heat kernel filter.

[0272] For the medium-speed diffusion heat kernel filter, the heat kernel filter coefficient The node feature update formula is:

[0273]

[0274] same, In this case, due to the characteristics of medium-speed diffusion, the update speed of node features is relatively moderate, which can timely reflect the changes in network status without causing instability due to too fast updates. Other features are also updated according to the corresponding rules.

[0275] For the slow diffusion heat kernel filter, the heat kernel filter coefficient The node feature update formula is:

[0276]

[0277] Here The slow diffusion type focuses on the deep propagation and long-term stability of features. The node features are updated relatively slowly and gradually adjusted to adapt to the long-term state change trend of the network to ensure the accuracy and stability of the data. The updates of other features are similar.

[0278] 3. Repeat the above state transfer and feature update process, and continue feature diffusion until the predetermined diffusion time is reached (which can be set according to the diffusion type of the heat kernel filter, for example, a shorter diffusion time can be set for the fast diffusion type, and a longer diffusion time can be set for the slow diffusion type) or certain convergence conditions are met (such as the change amplitude of the network state is less than a certain threshold, and the setting of the threshold can also be adjusted according to the type and parameters of the heat kernel filter, for example, a relatively large threshold can be set for the fast diffusion type, and a smaller threshold can be set for the slow diffusion type), indicating that the feature diffusion has stabilized.

[0279] After feature diffusion is completed, the diffused node feature matrix H′ can be obtained.

[0280] Step 3: Dynamically adjust the node features through a multi-layer perceptron to gradually approach the optimal solution.

[0281] The multilayer perceptron in this solution is a feedforward neural network consisting of multiple hidden layers that can achieve complex nonlinear mapping. MLP improves the model's expressiveness and generalization capabilities by learning nonlinear combinations of features layer by layer.

[0282] The diffused node feature matrix H′ is input into the above multilayer perceptron, and a new node feature representation H is generated through nonlinear transformation and feature combination. L .

[0283] In order to ensure that the feature update process can effectively reduce task delays and resource usage costs, these factors will be taken into consideration when setting the loss function and optimized through the gradient descent method.

[0284] Step 4: Flatten the multi-dimensional node features into a one-dimensional vector and concatenate them with the feature dimension identifier embedding.

[0285] In this step, the new multidimensional node feature H obtained in step 3 will be L It is flattened into a one-dimensional vector and concatenated with the feature dimension identifier embedding to simplify the feature processing process and improve computational efficiency while preserving the relationship between features to ensure the effectiveness of the optimization process.

[0286] This step will convert the input features into low-dimensional representations through the linear transformation layer, and convert the output features back to the output dimension. After the input features are processed by the linear transformation layer for dimensionality reduction, an intermediate result of low-dimensional representation is obtained. However, subsequent steps (such as feature concatenation and other operations) expect the features to have a specific dimensional format for efficient calculation and processing. Therefore, it is necessary to convert the low-dimensional output features obtained through linear transformation back to output dimensions suitable for processing in subsequent steps to ensure the consistency and compatibility of the data throughout the process. The role of this step is to achieve efficient feature mapping through linear transformation and ensure the consistency of input and output feature dimensions.

[0287] The multi-dimensional node features are flattened into one-dimensional vectors to facilitate subsequent linear transformation and processing. The flattened node features are represented as H flat , then each node feature h i Expand to a one-dimensional vector:

[0288]

[0289] In order to preserve the relationship between features and the semantic information of node features, feature dimension identifier embedding is introduced to concatenate the feature dimension identifier with the flattened feature vector to form an extended feature matrix H ext :

[0290] H ext =[H flat &D embed ]

[0291] Among them, D embed It is a feature dimension identification matrix, which is used to represent the semantic information of the feature.

[0292] In this step, the feature dimension identification matrix D is introduced embed The purpose is to preserve the relationship and semantic information between features to enhance the effect of feature processing and improve the accuracy of feature mapping.

[0293] The following is the detailed determination process of the feature dimension identification matrix:

[0294] Feature dimension identification matrix D embed is an embedding matrix that represents the semantic information of each feature dimension. For each feature dimension, we assign an embedding vector to it. These embedding vectors can capture the relationship between features and provide useful information during feature processing.

[0295] The feature dimension identification matrix can be determined by the following steps:

[0296] Assume that the feature dimension of the initial node feature matrix H is d, that is, each node feature vector i Contains d features. Set the embedding vector dimension k of the feature dimension identifier. The embedding vector dimension k is determined according to the specific task and computing resources, and is usually a small integer (such as 32, 64, 128, etc.).

[0297] Assign an embedding vector to each feature dimension. These embedding vectors can be randomly initialized or obtained through pre-training. Initial feature dimension identification matrix D embed is a d×k matrix, where each row represents an embedding vector of a feature dimension:

[0298]

[0299] Among them, d embedi Represents the embedding vector of the i-th feature dimension.

[0300] The embedding vector can be optimized through the training process to better represent the semantic information of the feature dimension. During the training process, the embedding vector will be updated according to the feedback of the loss function and gradually adjusted to the optimal state.

[0301] For example, suppose there is the following initial node feature matrix H:

[0302]

[0303] Among them, the number of feature dimensions d=4 (ie, computing capacity CC, storage capacity SC, current load CL, location L).

[0304] Set the embedding vector dimension k = 4. Assign an embedding vector to each feature dimension and initialize it randomly:

[0305]

[0306] Concatenate the embedding vector with the node feature matrix to construct the extended feature matrix H ext , concatenate the embedding vector with the node feature matrix H:

[0307]

[0308] During model training, the feature dimension identification matrix D embed It will be optimized based on the feedback of the loss function. The loss function L that comprehensively considers feature processing and task requirements can be set as:

[0309] L = l feature +L task

[0310] Among them, l feature represents the loss of feature processing; L task Represents a loss in mission demand.

[0311] Use gradient descent to optimize the loss function and update the embedding vector D embed :

[0312]

[0313] in, is the learning rate, is the gradient of the loss function with respect to the embedding vector.

[0314] In each iteration, the gradient of the loss function is calculated through the back-propagation algorithm and the embedding vector D is updated. embed .

[0315] The feature dimension identification matrix can retain the relationship and semantic information between features, enhance the effect of feature processing, accurately represent the semantic information of feature dimensions, and improve the accuracy of feature mapping. In addition, the feature dimension identification matrix can be dynamically adjusted according to task requirements and network environment to adapt to different optimization goals, thereby ensuring the efficiency and stability of feature processing and optimization process. By optimizing the feature dimension identification matrix, the accuracy of feature mapping can be significantly improved, providing a solid foundation for subsequent resource allocation and task scheduling.

[0316] Step 5: Develop the best resource allocation and task scheduling strategy through the optimized node feature representation.

[0317] The resource allocation strategy aims to reasonably allocate computing resources based on the node feature representation optimized by the above steps to ensure that the task can be completed efficiently.

[0318] Input the optimized node feature matrix H ext and task collection Each task t k Including computing requirements C k and delay requirement D k .

[0319] The goal of resource allocation is to minimize the task delay and resource usage cost and meet the task's computing needs and latency requirements.

[0320] After considering task delay and resource usage cost, the resource allocation objective function L is set alloc ,:

[0321]

[0322] Among them, α and β are weight coefficients, which can be set by relevant personnel according to experience and scenarios. For example, in some application scenarios with high real-time requirements, α and β can be 0.7 and 0.3 respectively; in some application scenarios with limited resources or cost control requirements, α and β can be 0.3 and 0.7 respectively. k ) and cost(t k ) represent tasks t k latency and resource usage costs.

[0323] Calculate each task t k Latency and resource usage costs:

[0324]

[0325] Among them, CC i Represents node v i The computing power of network_latency(t k,v i ) represents task t k From user device to node v i The network delay of compute_cost(C k ,v i ) and storage_cost(t k ,v i ) represent the computation and storage costs respectively.

[0326] Select the best resource allocation scheme so that the objective function L alloc Minimize, the calculation formula is as follows:

[0327]

[0328] Among them, allocation(t k ,v i ) indicates that the task t k Assign to node v i The optimal strategy.

[0329] The resource allocation process is as follows:

[0330] First, for each task t k , calculate its i latency and resource usage costs.

[0331] Then, all possible allocation schemes are compared and the one that minimizes the objective function is selected.

[0332] Then, the task t k Assign to the selected node v i .

[0333] When determining the task scheduling strategy, it is necessary to determine the execution order of tasks to ensure that tasks can be completed on time and maximize resource utilization.

[0334] Enter the above resource allocation plan and task set At this time, the task scheduling goal is to minimize the total completion time of the task and maximize the utilization of computing resources.

[0335] Therefore, the task scheduling objective function L can be sched Set to:

[0336]

[0337] Among them, start_time(t k ) represents task t k The start time of execution_time(t k ) represents task tk Execution time.

[0338] Then, calculate the start time and execution time of each task:

[0339]

[0340]

[0341] Among them, arrival_time(t k ) represents task t k start time; end_time(t k-1 ) represents the previous task t k-1 End time of CC i Indicates the node v to which the task is assigned i computing power.

[0342] Then, the best task scheduling order is selected so that the objective function L sched Minimize:

[0343]

[0344] The resource scheduling process is:

[0345] First, for each task t k , calculate its possible start time and execution time.

[0346] Then, all possible scheduling sequences are compared and the one that minimizes the objective function is selected.

[0347] Then, determine the task t k The start and end times of the task are updated.

[0348] In practical applications, network environments and task requirements change dynamically. Therefore, dynamic adjustment and real-time optimization are required to ensure the efficiency of resource allocation and task scheduling. For example, according to real-time monitoring data, resource allocation and task scheduling strategies are dynamically adjusted. When changes in network status or task requirements are detected, resource allocation and task scheduling schemes are recalculated.

[0349] In this scheme, we can first use graph structure modeling to make full use of network topology information to provide a basis for subsequent feature diffusion and optimization, and use hot kernel filters to diffuse and smooth node features to generate a diffused feature matrix to ensure that negative feedback is not introduced during feature diffusion and avoid instability in the optimization process. Then, through the nonlinear mapping capability of the multilayer perceptron, the feature representation is dynamically adjusted to gradually approach the optimal solution, reducing task delays and resource usage costs. The obtained multidimensional features are flattened into one-dimensional vectors, and a feature dimension identification matrix is ​​introduced to retain the relationship between features and semantic information, which can improve the efficiency of feature processing and the effect of feature processing. Finally, according to the optimized feature representation, computing resources are reasonably allocated and task scheduling strategies are set.

[0350] Through the organic combination of the above five key steps, this solution can effectively attract the optimization trajectory to reach the deep local minimum in a complex network environment, thereby achieving better resource allocation and task scheduling. The meticulous design and mutual coordination of each step ensure the efficiency and stability of the optimization process, providing strong technical support for 5G edge computing.

[0351] Compared with the existing technology, the solution of this application first constructs a graph structure of edge computing nodes and user devices and initializes features to provide a basic data framework for subsequent operations; then uses a hot kernel filter to classify according to application scenarios and combines Markov chains for feature diffusion to stably and accurately propagate features and optimize strategies; then dynamically adjusts node features through a multi-layer perceptron to improve model capabilities and approach the optimal solution; then flattens multi-dimensional node features and embeds feature dimension identifiers to simplify the processing process and retain feature relationships; finally, formulates resource allocation and task scheduling strategies based on the optimized features, while taking into account factors such as task delay and cost, to achieve better resource allocation and task scheduling in complex network environments, effectively improving edge computing resource utilization, reducing network delays and improving overall service quality.

[0352] In addition, in terms of feature diffusion, the present application scheme introduces a Markov chain, classifies the heat kernel filters according to different application scenarios and sets different time parameters, determines the transfer probability matrix based on the node state characteristics, heat kernel filter characteristics and network topology, so that the feature diffusion process is more accurate and stable, and can better reflect the network dynamics. By setting the transfer probability matrix in combination with the characteristics of different types of heat kernel filters, the diffusion process of node features in the network can be more accurately simulated. The advantages of different heat kernel filters in different network scenarios are fully utilized, making feature propagation more in line with the actual network dynamic change law. Considering the influence of heat kernel filter parameters on the transfer probability, the accuracy of feature diffusion simulation is further improved. Based on more accurate feature diffusion simulation, it can better capture the complex relationships between nodes and the details of feature propagation, providing a more reliable basis for resource allocation and task scheduling.

[0353] In addition, the present application scheme combines the stability of the heat kernel filter and the nonlinear expression ability of the multilayer perceptron, effectively attracting the optimization trajectory to reach the deep local minimum in a complex network environment, achieving better resource allocation and task scheduling, thereby significantly improving the utilization of edge computing resources.

[0354] Moreover, the nonlinear mapping capability of the multi-layer perceptron is utilized to dynamically adjust the node feature representation so that it gradually approaches the optimal solution, thus improving the accuracy and rationality of the feature representation. In addition, the multi-dimensional features are flattened into one-dimensional vectors, and the feature dimension identification matrix is ​​introduced to retain the relationship and semantic information between the features. Efficient feature mapping is achieved through the linear transformation layer, which simplifies the calculation process and improves the efficiency and accuracy of feature processing.

[0355] In this scheme, the Markov chain is creatively introduced into the feature diffusion process, and the heat kernel filter is innovatively classified and different time parameters are set according to different application scenarios to determine the transition probability matrix based on the node state characteristics, heat kernel filter characteristics and network topology structure, so that the feature diffusion process is more accurate and stable, and can better reflect the network dynamics. By setting the transition probability matrix in combination with the characteristics of different types of heat kernel filters, the diffusion process of node features in the network can be more accurately simulated. The advantages of different heat kernel filters in different network scenarios are fully utilized, making the feature propagation more in line with the actual network dynamic change law. Considering the influence of heat kernel filter parameters on the transition probability, the accuracy of feature diffusion simulation is further improved. Based on more accurate feature diffusion simulation, the complex relationships between nodes and the details of feature propagation can be better captured, providing a more reliable basis for resource allocation and task scheduling.

[0356] Therefore, the present application scheme can not only achieve better resource allocation and task scheduling, significantly improve the utilization rate of edge computing resources, but also make the feature diffusion process more accurate, improve the accuracy and rationality of feature representation, simplify the calculation process, and improve the efficiency and accuracy of feature processing.

[0357] This application scheme combines the nonlinear mapping capabilities of graph neural networks and multi-layer perceptrons to achieve efficient allocation and task scheduling of 5G edge computing resources, significantly improve resource utilization and service quality, reduce computing delays, reduce operating costs, enhance system stability and adaptability, and support large-scale deployment. This solution has a wide range of commercial application value and can achieve efficient, stable and intelligent system operations in multiple fields such as smart cities, industrial Internet of Things, intelligent transportation systems, intelligent manufacturing, smart medical care, smart homes, content distribution networks (CDNs), cloud computing and data centers, unmanned driving and Internet of Vehicles, promote industry development and technological innovation, improve user experience, and meet diverse business needs.

[0358] This application also provides an edge computing resource processing device, please refer to Figure 4 , the edge computing resource processing device includes:

[0359] A construction module 10 is used to construct a graph structure of edge computing nodes and user devices, and initialize the features of the graph structure to obtain node features of the graph structure;

[0360] A diffusion module 20, used to diffuse the node features of the graph structure to obtain a node feature matrix after diffusion;

[0361] An adjustment module 30, used for inputting the diffused node feature matrix into a multi-layer perceptron to dynamically adjust the node features and generate a new node feature representation;

[0362] A concatenation module 40 is used to concatenate the new node feature representation with a predetermined feature dimension identification matrix to generate an extended feature matrix;

[0363] The resource allocation and scheduling module 50 is used to perform resource allocation and scheduling according to the extended feature matrix.

[0364] The edge computing resource processing device provided in the present application adopts the edge computing resource processing method in the above-mentioned embodiment to achieve better resource allocation and task scheduling in a complex network environment, significantly improving the utilization rate of edge computing resources and resource scheduling efficiency. Compared with the prior art, the beneficial effects of the edge computing resource processing device provided in the present application are the same as the beneficial effects of the edge computing resource processing method provided in the above-mentioned embodiment, and other technical features in the edge computing resource processing device are the same as the features disclosed in the above-mentioned embodiment method, which will not be repeated here.

[0365] The present application provides an edge computing resource processing device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the edge computing resource processing method in the above-mentioned embodiment one.

[0366] Reference below Figure 5 , which shows a schematic diagram of the structure of the edge computing resource processing device used to implement the embodiment of the present application. The edge computing resource processing device in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 5 The edge computing resource processing device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0367] like Figure 5As shown, the edge computing resource processing device may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the edge computing resource processing device are also stored. The processing device 1001, ROM1002, and RAM1004 are connected to each other via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD: Liquid Crystal Display), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the edge computing resource processing device to communicate wirelessly or wired with other devices to exchange data. Although the figure shows an edge computing resource processing device with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be implemented or provided alternatively.

[0368] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0369] The edge computing resource processing device provided by this application adopts the edge computing resource processing method in the above embodiment to solve the technical problems of edge computing resource processing. Compared with the prior art, the beneficial effects of the edge computing resource processing device provided by this application are the same as the beneficial effects of the edge computing resource processing method provided by the above embodiment, and the other technical features in the edge computing resource processing device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0370] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0371] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0372] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer programs) stored thereon, and the computer-readable program instructions are used to execute the edge computing resource processing method in the above-mentioned embodiment.

[0373] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0374] The above-mentioned computer-readable storage medium may be included in the edge computing resource processing device; or it may exist independently without being assembled into the edge computing resource processing device.

[0375] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by the edge computing resource processing device, the edge computing resource processing device: constructs a graph structure of edge computing nodes and user devices, and initializes the features of the graph structure to obtain the node features of the graph structure; diffuses the node features of the graph structure to obtain a diffused node feature matrix; inputs the diffused node feature matrix into a multilayer perceptron to dynamically adjust the node features to generate a new node feature representation; splices the new node feature representation with a predetermined feature dimension identification matrix to generate an extended feature matrix; and performs resource allocation and scheduling according to the extended feature matrix.

[0376] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0377] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0378] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0379] The readable storage medium provided in the present application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned edge computing resource processing method, thereby achieving better resource allocation and task scheduling in a complex network environment, and significantly improving the utilization rate of edge computing resources and resource scheduling efficiency. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in the present application are the same as the beneficial effects of the edge computing resource processing method provided in the above-mentioned embodiments, and will not be repeated here.

[0380] The present application also provides a computer program product, including a computer program, which implements the steps of the edge computing resource processing method as described above when executed by a processor.

[0381] The computer program product provided in the present application realizes better resource allocation and task scheduling in a complex network environment, significantly improving the utilization rate of edge computing resources and resource scheduling efficiency. Compared with the prior art, the beneficial effects of the computer program product provided in the present application are the same as the beneficial effects of the edge computing resource processing method provided in the above-mentioned embodiment, which will not be repeated here.

[0382] One or more technical solutions proposed in this application have at least the following technical effects: by constructing a graph structure of edge computing nodes and user devices and initializing features, a basic data framework is provided for subsequent operations; a hot kernel filter is used to classify application scenarios and combined with a Markov chain for feature diffusion to stably and accurately propagate features and optimize strategies; node features are then dynamically adjusted through a multi-layer perceptron to improve model capabilities and approach the optimal solution; multi-dimensional node features are flattened and embedded with feature dimension identifiers to simplify the processing process and retain feature relationships; finally, resource allocation and task scheduling strategies are formulated based on the optimized features, while taking into account factors such as task delay and cost, to achieve better resource allocation and task scheduling in complex network environments, effectively improving edge computing resource utilization, reducing network delays, and improving overall service quality.

[0383] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A method for processing edge computing resources, characterized in that: The method comprises: Constructing a graph structure of edge computing nodes and user devices, and initializing features of the graph structure to obtain node features of the graph structure; Diffusion of node features of the graph structure is performed to obtain a diffused node feature matrix; Inputting the diffused node feature matrix into a multi-layer perceptron to dynamically adjust the node features and generate a new node feature representation; The new node feature representation is concatenated with a predetermined feature dimension identification matrix to generate an extended feature matrix; Resource allocation and scheduling are performed according to the extended feature matrix.

2. The method according to claim 1, characterized in that The step of diffusing the node features of the graph structure to obtain a diffused node feature matrix comprises: The node features of the graph structure are diffused by using a heat kernel filter according to application scenario classification and combined with a Markov chain to obtain a diffused node feature matrix.

3. The method according to claim 2, characterized in that The step of using the heat kernel filter to classify the application scenarios and diffuse the node features of the graph structure in combination with the Markov chain to obtain a diffused node feature matrix includes: Combined with different application scenarios of edge computing networks, heat kernel filters are classified to obtain the diffusion type of the heat kernel filters; According to the diffusion type of the heat kernel filter, setting corresponding time parameters for the heat kernel filter; Determine the corresponding heat kernel filter coefficient according to the time parameter of the heat kernel filter; Determining a transfer probability matrix of the heat kernel filter based on the heat kernel filter coefficients and in combination with a Markov chain; The node features of the graph structure are diffused according to the transition probability matrix to obtain a diffused node feature matrix.

4. The method according to claim 3, characterized in that The step of determining the transfer probability matrix of the heat kernel filter based on the heat kernel filter coefficient and in combination with the Markov chain comprises: Determine a state vector of each node according to the node characteristics of the graph structure; Construct the state space of the Markov chain based on the state vector of each node; The transfer probability matrix of the heat kernel filter is calculated based on the heat kernel filter coefficients and in combination with the state space and network topology of the Markov chain.

5. The method according to claim 4, characterized in that The step of calculating the transfer probability matrix of the heat kernel filter according to the heat kernel filter coefficient and in combination with the state space and network topology structure of the Markov chain comprises: Calculating the state influence weight of the neighboring nodes of the target node based on the state vectors of the target node and the neighboring nodes in the graph structure; Determining the influence weight of the thermal kernel filter parameters according to the thermal kernel filter coefficients; Determine the node state maintenance weight of the target node; The transfer probability of the target node is calculated according to the adjacent node state influence weight, the heat kernel filter parameter influence weight and the node's own state retention weight; Based on the state space of the Markov chain and the transition probability of each target node in the graph structure, a transition probability matrix of the heat kernel filter is calculated.

6. The method according to claim 3, characterized in that The step of initializing the features of the graph structure and obtaining the node features of the graph structure comprises: Initializing node features of the graph structure to obtain a node feature vector; The initial node feature matrix is ​​obtained by constructing a matrix based on the node feature vectors of each node; The step of diffusing the node features of the graph structure according to the transition probability matrix to obtain a diffused node feature matrix comprises: The initial node feature matrix is ​​updated according to the transition probability matrix to obtain a diffused node feature matrix.

7. The method according to claim 6, characterized in that The step of determining the characteristic dimension identification matrix comprises: Determining the characteristic dimension of the initial node characteristic matrix; Assign an embedding vector to each feature dimension; Based on the embedding vector and the embedding vector dimensions of the predetermined feature dimension identifiers, an initial feature dimension identifier matrix is ​​obtained; Training and optimizing the embedding vector in combination with the initial node feature matrix to obtain an optimized embedding vector; An optimized feature dimension identification matrix is ​​determined based on the optimized embedding vector.

8. The method according to any one of claims 1 to 7, characterized in that The step of allocating resources according to the extended feature matrix comprises: Determine a task set, each task in the task set includes a computing requirement and a latency requirement; Calculate the latency and resource usage cost of each task on each node; Setting a resource allocation objective function according to a resource allocation goal, wherein the resource allocation goal is to minimize the delay and resource usage cost of the task and to meet the computing requirements and latency requirements of the task; Determining an optimal resource allocation scheme according to the extended characteristic matrix to minimize the resource allocation objective function; The corresponding tasks are allocated to the corresponding nodes according to the optimal resource allocation scheme.

9. The method according to claim 8, characterized in that The step of performing resource scheduling according to the extended feature matrix comprises: Setting a task scheduling objective function according to the task scheduling goal, wherein the task scheduling goal is to minimize the total completion time of the task and maximize the utilization rate of computing resources; Calculate the start time and execution time of each task; According to the start time and execution time of each task, determine the execution order of each task to minimize the task scheduling objective function; Resource scheduling is performed according to the determined execution order of each task.

10. An edge computing resource processing device, characterized in that: The edge computing resource processing device comprises: A construction module is used to construct a graph structure of edge computing nodes and user devices, and initialize the features of the graph structure to obtain node features of the graph structure; A diffusion module, used for diffusing the node features of the graph structure to obtain a node feature matrix after diffusion; An adjustment module, used for inputting the diffused node feature matrix into a multi-layer perceptron to dynamically adjust the node features and generate a new node feature representation; A concatenation module, used for concatenating the new node feature representation with a predetermined feature dimension identification matrix to generate an extended feature matrix; The resource allocation and scheduling module is used to perform resource allocation and scheduling according to the extended feature matrix.

11. An edge computing resource processing device, characterized in that: The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the edge computing resource processing method according to any one of claims 1 to 10.

12. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the steps of the edge computing resource processing method according to any one of claims 1 to 9.

13. A computer program product, characterized in that The computer program product includes a computer program, and when the computer program is executed by a processor, it implements the edge computing resource processing method according to any one of claims 1 to 9.

Images