A network sparsification measurement method and system based on a graph neural network

Abstract

The application relates to the technical field of network sparsification measurement, and provides a network sparsification measurement method based on a graph neural network, which comprises the following steps: S1: performing sparsification sampling on end-to-end network service quality data; S2: based on the sparsification sampled data, establishing a network service quality prediction model and training; S3: in a manner of a continuous learning function, using historically learned space-time network service quality information for learning and reasoning at a future moment. The technical scheme has fully tapped the space-time correlation of edge cloud network service quality, and uses the space-time correlation between network service qualities to complete a full-amount network end-to-end service quality matrix. The technical scheme solves the problem that, in a traditional algorithm, while filtering out abnormal data, part of useful data is filtered out, can cope with complex and changeable network service quality data, and solves the problem that, in a traditional algorithm, continuous reasoning cannot be performed during completion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of network sparsification measurement, and in particular to a network sparsification measurement method and system based on a graph neural network. BACKGROUND

[0002] In recent years, the demand for distributed systems has been increasing. The emergence of more and more delay-sensitive applications, such as VR, AR, and autonomous driving, has led service providers to increasingly prefer dense and distributed deployment of services. Edge cloud, as an emerging edge computing mode, allows dense and distributed deployment of services, and is attracting people's attention. In the face of such a distributed edge cloud system, service providers need to know all the end-to-end network quality of service (QoS) in order to make appropriate resource arrangements and requests. In addition, when a distributed system fails, comprehensive and accurate end-to-end service quality can better help service providers estimate the number of affected users and solve the problem of failure.

[0003] However, the dense deployment of services poses a huge challenge to the monitoring of network QoS. For example, for an edge cloud system with n nodes, a full end-to-end QoS detection requires n 2 level measurements, which puts an unbearable, huge burden on the system. Even without considering these burdens, network errors, data transmission delays, etc. also bring strong sparseness to the full QoS. Therefore, the present application aims to complete the real measured, sparse end-to-end QoS through a data completion algorithm. In this way, we can reduce the complexity of measurement, improve the efficiency of measurement, and enhance the control of network QoS.

[0004] Existing network service quality completion algorithms are mostly matrix completion algorithms based on mathematics, which cannot cope with the complexity and variability of edge cloud data. At the same time, they have the following disadvantages:

[0005] (1) The complexity and variability of edge cloud service quality are not considered. The matrix is decomposed and then completed through mathematical methods. This method filters out some abnormal data, but also weakens useful features in the data, affecting the prediction accuracy.

[0006] (2) The role of existing learned patterns in future data completion is not considered. Existing applications such as VR, AR, and autonomous driving have high real-time requirements for network services. This requires network service providers to also have real-time awareness of network service quality. Mathematical methods retrain and reason at each completion, and cannot apply existing learned completion patterns to new data reasoning.

[0007] (3) Not considering the help of space-time relationship for service quality data reasoning. Network service quality data has strong space-time correlation. In particular, the end-to-end delay between two servers has a strong correlation with the physical distance between the two servers. The traditional mathematical method cannot effectively utilize the space-time correlation between nodes, thereby affecting the accuracy of reasoning and the calculation efficiency of reasoning. SUMMARY

[0008] In view of the above problems, the purpose of the present application is to provide a network sparsification measurement method and system based on graph neural network, which is used to solve the problems mentioned in the above background technology.

[0009] The above invention purpose of the present application is realized by the following technical scheme:

[0010] A network sparsification measurement method based on graph neural network, comprising the following steps:

[0011] S1: Sparsification sampling of end-to-end network service quality data;

[0012] S2: Based on the sparsification sampled data, a network service quality prediction model is established and trained;

[0013] S3: Using the way of continuous learning function, the historical learned space-time network service quality information is used for future learning and reasoning.

[0014] Further, the end-to-end network service quality data includes end-to-end delay data, end-to-end bandwidth data and end-to-end packet loss rate data.

[0015] Further, in step S1, the end-to-end network service quality data is sparsification sampled, specifically:

[0016] S11: Formulate a network sparsification measurement plan, for a distributed system containing n network nodes, containing n 2 end-to-end paths, in all the end-to-end paths, according to the sampling rate alpha to randomly sample, wherein the random sampling is sampled in the way of Bernoulli sampling;

[0017] S12: Execute the network sparsification measurement plan, and save the sampling result of the network sparsification measurement plan to the center node.

[0018] Further, in step S2, the network service quality prediction model is established and trained, specifically:

[0019] S21: The delay matrix of the end-to-end path is constructed as graph structure data;

[0020] S22: a feature encoder based on a graph neural network, which encodes the features shown by the graph structure data by using a multi-layer encoding layer based on message passing;

[0021] S23: a delay predictor based on a multi-layer perceptron, which restores the hidden features generated by encoding the graph structure data in step S22 into a complete network service quality prediction matrix;

[0022] S24: a gradient descent method is used for repeated training and optimization to minimize the error.

[0023] Further, in step S21, the delay matrix of the end-to-end path is constructed as the graph structure data, specifically:

[0024] The distributed system containing n network nodes constructs the graph structure data as G(V, E), where V = {v1, v2,... vn} is the set of all n network nodes, e n} represents the network service quality from network node i to network node j. ij

[0025] Further, in step S22, the feature encoder based on the graph neural network encodes the features shown by the graph structure data by using a multi-layer encoding layer based on message passing, specifically:

[0026] For any u∈V, the network service quality information between the neighbor network nodes of u is embedded into the representation of u by aggregation, and the aggregation formula is:

[0027]

[0028] where AGG l is an aggregation function, σ is a nonlinear function sigmod, CONCAT represents concatenation of two vectors, l is a layer number, v is a node to be encoded, is a learnable parameter for encoding in the lth layer, is the feature representation of node x in the lth layer of node u, is the feature representation of the service quality from node u to node v in the (l-1)th layer;

[0029] After aggregation, the current layer aggregation result is combined with the information of the previous layer v representation to obtain the output of the current layer The expression is:

[0030]

[0031] where, ​The learnable parameter matrix for the lth layer for aggregation, The output of the l-1th layer is used to initialize h0 using the feature x of v when the current layer is the first layer;

[0032] The inter-layer residual is defined, and the inter-layer residual is injected into the output y of the lth layer l Specifically, the information of the previous layer is input to the current layer at a ratio of 1-λ:

[0033]

[0034] Wherein, The output of the previous layer;

[0035] The initial residual is defined to compensate for the gradient vanishing and gradient explosion caused by multi-layer graph convolution coding, and the expression is:

[0036]

[0037] Wherein, x0 is the initial feature of all nodes, and β is the adaptive scaling ratio determined by the Frobenius norm;

[0038]

[0039] Wherein, h l The feature matrix of all node features of the lth layer is spliced, and x0 is the initial feature matrix of all nodes;

[0040] Finally, the inter-layer residual and the initial output are superimposed on the output of the graph neural network Output on the above, the expression is:

[0041]

[0042] At this time, the encoding of the features shown by the graph structure data of the lth layer is completed.

[0043] Further, in step S23, the hidden features generated by encoding the graph structure data in step S22 are restored into a complete network service quality prediction matrix using a multi-layer perceptron-based delay predictor, specifically:

[0044] In step S2, for any network node v, the hidden feature

[0045] The service quality matrix is reconstructed using a multi-layer perceptron-based decoding method to restore the high-dimensional hidden representation of the service quality matrix, and the expression is:

[0046]

[0047] in, The final service quality prediction output of the model is given by t, where t is the current time, W is the learnable parameter matrix used for the output, T is the transpose of the matrix, and b is the bias coefficient matrix.

[0048] Furthermore, in step S24, the gradient descent method is used to minimize the error, specifically as follows:

[0049] The mean squared error is used to describe the difference between predicted and actual values, and its expression is:

[0050]

[0051] Where l represents the error between the predicted value and the actual value. To predict service quality, A t For genuine service quality;

[0052] Based on the obtained error, gradient descent is used for repeated training and optimization. During the training process, the Adam optimizer is used to dynamically adjust the learning rate until the error change between two optimizations is less than a preset threshold.

[0053] Furthermore, in step S3, a continuous learning function is used to apply the spatiotemporal network service quality information learned in the past to the learning and inference of future moments, specifically as follows:

[0054] F′ n+1 =G(F0, F1, ..., F n )

[0055] Where G is the continuous learning function, and the input is the historical models F0, F1...F n The output is the initial model F′ at time n+1. n+1 At time n+1, first use F′ n+1 Initialize the model and train it based on the initialization.

[0056] A graph neural network-based network sparsity measurement system for performing the above-described graph neural network-based network sparsity measurement method includes:

[0057] The sparse data acquisition module is used to perform sparse sampling of end-to-end network service quality data.

[0058] The model building and training module is used to build and train a network service quality prediction model based on the sparsely sampled data.

[0059] The continuous learning module is used to apply historically learned spatiotemporal network service quality information to future learning and inference using continuous learning functions.

[0060] A computer device comprising a memory and one or more processors, the memory having stored therein computer code that, when executed by the one or more processors, causes the one or more processors to perform the method as described above.

[0061] A computer-readable storage medium having stored therein computer code that, when executed, performs the method as described above.

[0062] Compared with the prior art, the present application includes at least one of the following beneficial effects:

[0063] (1) Fully tap the spatio-temporal correlation of edge cloud network service quality, and use the spatio-temporal correlation between network service qualities to complete the full-quantity network end-to-end service quality matrix.

[0064] (2) Solve the problem that the traditional algorithm filters out part of the useful data while filtering out the abnormal data, and can cope with complex and variable network service quality data.

[0065] (3) Solve the problem that the traditional algorithm cannot perform continuous reasoning when completing. The present application can apply existing and historical data patterns to new data reasoning, reduce the computational overhead caused by training, and enable the entire framework to perform fast and continuous stable reasoning. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 The overall flowchart of the network sparsification measurement method based on the graph neural network of the present application;

[0067] Figure 2 The overall structure diagram of the network sparsification measurement system based on the graph neural network of the present application. DETAILED DESCRIPTION

[0068] To make the purposes, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0069] As will be understood by one of ordinary skill in the art, the singular forms "a," "an," "said," and "the" include plural referents unless the context clearly dictates otherwise. It should be further understood that the terms "comprise," "comprises," "comprising," "include," "includes," "including," "contain," "contains," or "containing," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, or groups thereof.

[0070] Conventional fault prediction is mainly based on supervised classification or prediction methods, and edge cloud scenarios have more complex scenarios and more uncontrollable factors.

[0071] The method mainly has the following technical breakthroughs:

[0072] (1) The patent provides an algorithm for completing network service quality data based on a graph neural network (GNN), which can fully utilize the spatio-temporal correlation between nodes to fill in the missing parts in the overall end-to-end service quality data.

[0073] (2) On the basis of using a graph neural network, the patent also designs a set of adaptive residual structures to compensate for the shortcomings that useful features are weakened when using a graph neural network. The use of residual structures can effectively extract high-frequency and valuable features from the original data.

[0074] (3) The patent uses a continuous learning method to reduce the computational cost of the neural network. The patent proposes a continuous learning method based on reasoning knowledge learned from existing data to reason about unknown parts of the service quality data. Using this method, we can reduce the number of iterations of the neural network, thereby reducing the computational complexity.

[0075] The terms involved in the present application are explained as follows:

[0076] Graph neural network:

[0077] Graph Neural Network (GNN) is a framework that uses deep learning to directly learn graph structure data in recent years. Its excellent performance has attracted the attention and in-depth exploration of scholars. By formulating certain strategies on the nodes and edges of the graph, GNN converts the graph structure data into a standard and standardized representation and inputs it into various neural networks for training, achieving excellent results in node classification, edge information propagation, and graph clustering tasks.

[0078] The history of GNN can be traced back to 2005, when Gori et al. first proposed the concept of GNN using RNN to process undirected graphs, directed graphs, labeled graphs, and cyclic graphs. Since then, Scarselli et al. and Micheli et al. have inherited and developed the GNN algorithm of this mode and made some improvements. The early stage of GNN is mainly based on RNN as the main framework, which generates vector expression for each node through simple feature mapping and node aggregation, and cannot well cope with the complex and variable graph data in reality. In view of this situation, Bruna et al. proposed to apply CNN to the graph, and through the ingenious conversion of convolution operator, they proposed the graph convolution network (GCN), and derived many variants. GCN realizes the translation invariance, local perception and weight sharing of CNN on the graph, and provides ideological guidance and reference for the construction and improvement of other GNN frameworks.

[0079] From the proposal of GNN concept by Gori et al. in 2005 to the emergence of GCN providing an effective processing paradigm for non-Euclidean structure data, to the proposal of different GNN framework variants such as GAE, GAT, GRN, GGN and the application of GNN in various fields, GNN has experienced a process from nothing to something and from something to optimization in theory and practice, and the GNN system family is also constantly developing and improving. From this process, we can see that many researchers are constantly improving and optimizing the GNN algorithm and structure.

[0080] Residual structure:

[0081] The characteristics of residual network are easy to optimize and can improve accuracy by increasing the depth. The residual block inside uses a jump connection to alleviate the gradient vanishing problem caused by increasing the depth in deep neural networks.

[0082] Continual learning:

[0083] Continual learning refers to a model that learns a large number of tasks in sequence without forgetting the knowledge obtained from previous tasks. This is an important concept because under the premise of supervised learning, machine learning models are trained as the best function for a given data set or data distribution. In the real environment, data is rarely static and may change. When faced with unseen data, typical ML models may perform poorly. This phenomenon is called catastrophic forgetting.

[0084] Bernoulli sampling:

[0085] Bernoulli sampling is a random sampling method based on Bernoulli distribution. In Bernoulli sampling, each event has two possible outcomes, such as success or failure, head or tail, etc. These outcomes are randomly generated according to a Bernoulli distribution with a parameter p, where p represents the probability of success. If an event is successful, we mark it as 1, otherwise as 0.

[0086] Bernoulli sampling is often used in machine learning and statistics, for example, in binary classification tasks, we can use Bernoulli sampling to generate sample data, where success represents positive examples and failure represents negative examples.

[0087] The implementation of Bernoulli sampling can use pseudo-random number generators, such as functions in the random module in Python. Specifically, we can use the random.random() function to generate a random number r between 0 and 1, and if r is less than or equal to p, we mark the event as 1, otherwise as 0. In this way, we can generate a sample dataset, where the outcome of each event is randomly generated according to the probability of p.

[0088] First embodiment

[0089] As Figure 1 shown, the present embodiment provides a network sparsification measurement method based on graph neural network, comprising the following steps:

[0090] S1: Sparsification sampling of end-to-end network service quality data.

[0091] S2: Based on the sparsification sampling data, establish network service quality prediction model and train.

[0092] S3: Using the way of continuous learning function, the historical learning of space-time network service quality information is used for future learning and reasoning.

[0093] The following describes each step of the present invention in detail:

[0094] In step S1, the end-to-end network service quality data is sparsified and sampled, specifically:

[0095] S11: Develop a network sparsification measurement plan, for a distributed system containing n network nodes, containing n 2 end-to-end paths, in all the end-to-end paths, according to the sampling rate α to randomly sample, wherein the random sampling is sampled in the way of Bernoulli sampling.

[0096] S12: Execute the network sparsification measurement plan, and save the sampling results of the network sparsification measurement plan to the center node.

[0097] The end-to-end network service quality data includes end-to-end delay data, end-to-end bandwidth data, and end-to-end packet loss rate data.

[0098] In step S2, a network service quality prediction model is established and trained, specifically:

[0099] S21: The delay matrix of the end-to-end path is constructed as graph structure data, specifically:

[0100] The graph structure data constructed by the distributed system containing n network nodes is G(V, E), where V = {v1, v2,... vn}, E = {e1, e2,... en}, and n is the number of network nodes. n} is the set of all n network nodes, e ij represents the network service quality from network node i to network node j.

[0101] S22: Based on the graph neural network feature encoder, the features exhibited by the graph structure data are encoded using a multi-layer encoding layer based on message passing, specifically:

[0102] In the first part of the graph neural network feature encoding, a multi-layer encoding layer based on message passing is used to encode the features. For any u ∈ V, we obtain the embedding representation of u by aggregating the attributes on all edges related to u, and embed the network service quality information between the neighbor network nodes of u into the representation of u by aggregation, and the aggregation formula is:

[0103]

[0104] where AGG l is the aggregation function, σ is the nonlinear function sigmod, CONCAT represents concatenation of two vectors, l is the layer number, v is the node to be encoded, is the learnable parameter for encoding in the lth layer, is the feature representation of node x in the lth layer of node u, is the feature representation of the service quality from node u to node v in the (l-1)th layer;

[0105] After aggregation, the current layer aggregation result is combined with the information of the previous layer v representation to obtain the output of the current layer The expression is:

[0106]

[0107] where, is the learnable parameter matrix for aggregation in the lth layer, Output for the l-1 layer, using the feature x of v to initialize h0 when the current layer is the first layer;

[0108] The above completes the node topology information coding based on the graph neural network. However, the aggregation mode based on the aggregation function AGG l may weaken the useful features of the node, so we propose an adaptive residual module to make up for the weakening of GNN for rapidly changing network features. In this patent, we define two residual paths in total, which respectively inject inter-layer residuals and initial residuals into the output y l of the l-th layer.

[0109] Define the inter-layer residual, and inject the inter-layer residual into the output y l of the l-th layer, specifically, input the information of the previous layer to the current layer at a ratio of 1-λ:

[0110]

[0111] wherein, is the output of the previous layer;

[0112] On the basis of the above information fusion, the encoding of multiple layers will lead to potential gradient disappearance and gradient explosion, so here we design another residual route to make up for the gradient disappearance and gradient explosion caused by multi-layer graph convolution coding. Define the initial residual to make up for the gradient disappearance and gradient explosion caused by multi-layer graph convolution coding, and the expression is:

[0113]

[0114] wherein, x0 is the initial feature of all nodes, and we need to scale x0 by β and then superimpose it on . However, the size of the l-th layer relative to x0 is not the same, so we set β to be an adaptive value to adapt to the value range of different layers, and β is an adaptive scaling ratio determined by the Frobenius norm.

[0115]

[0116] wherein, h l is a feature matrix formed by concatenating the features of all nodes in the l-th layer, and x0 is an initial feature matrix of all nodes;

[0117] Finally, the inter-layer residual and the initial output are superimposed on the output of the graph neural network , and the expression is:

[0118]

[0119] At this time, the encoding of the features shown by the graph structure data of the first layer is completed.

[0120] S23: The hidden features generated by encoding the graph structure data in step S22 are restored into a complete network service quality prediction matrix using a multilayer perceptron-based delay predictor, specifically:

[0121] The second part is a multilayer perceptron-based delay predictor. In step S2, for any network node v, the hidden features

[0122] In this step, a multilayer perceptron-based decoder is used to reconstruct the service quality matrix, reducing the high-dimensional hidden representation of the service quality matrix. The expression is:

[0123]

[0124] wherein, is the final model service quality prediction output, t is the current time, W is the learnable parameter matrix for output, T is the transpose operation of the matrix, and b is the bias coefficient matrix.

[0125] Through the above decoder, the high-dimensional hidden representation of the service quality matrix is restored.

[0126] S24: Gradient descent method is used for repeated training and optimization to minimize the error, specifically:

[0127] The mean square error is used to describe the difference between the predicted value and the true value, and the expression is:

[0128]

[0129] wherein, l is the error between the predicted value and the true value, is the predicted service quality, A t is the true service quality.

[0130] Based on the obtained error, the gradient descent method is used for repeated training and optimization. In the training process, the Adam optimizer is used to dynamically adjust the learning rate until the error change between two optimizations is less than a preset threshold (such as 0.001).

[0131] In step S3, the historical learned spatio-temporal network service quality information is used for future learning and reasoning using a continuous learning function, specifically:

[0132] Since training and inference from scratch for the model will take a long time, and the implicit spatio-temporal correlation between the nodes in the distributed system does not change much in a short time, we process the historical inference matrix through continuous learning, and use the historical learned spatio-temporal network service quality information for future learning and inference, so as to reduce the computational complexity of training and inference for new time points.

[0133] F' n+1 =G(F0,F1,...,Fn) n )

[0134] Where G is a continuous learning function, the input is the historical model F0, F1...Fn n , and the output is the initial model F' of n+1 time n+1 , the model is initialized at the n+1 time using F' n+1 , and the model is trained based on the initialization.

[0135] Second embodiment

[0136] As shown in Figure 2 , the embodiment provides a graph neural network-based network sparsification measurement system for performing the graph neural network-based network sparsification measurement method as in the first embodiment, comprising:

[0137] A sparse data acquisition module 1 for sparse sampling of end-to-end network service quality data

[0138] A model establishment and training module 2 for establishing a network service quality prediction model and training based on the sparse sampled data.

[0139] A continuous learning module 3 for using a continuous learning function to use the historical learned spatio-temporal network service quality information for future learning and inference.

[0140] A computer readable storage medium, the computer readable storage medium stores computer code, when the computer code is executed, the above method is executed. Those skilled in the art can understand that all or part of the steps in the above embodiments can be completed by programs instructing related hardware, and the programs can be stored in a computer readable storage medium, and the storage medium can include read only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc.

[0141] The above merely describes the preferred embodiments of the present application, and the protection scope of the present application is not limited to the above-described embodiments. Any technical scheme falling within the concept of the present application shall be considered as falling within the protection scope of the present application. It should be noted that, for ordinary skilled in the art, some improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements shall be considered as falling within the protection scope of the present application.

[0142] The technical features of the above-described embodiments can be combined arbitrarily. In order to make the description simple, all possible combinations of the technical features in the above-described embodiments are not described, however, as long as the combinations of the technical features do not contradict each other, they shall be considered as falling within the scope of the present application.

[0143] It should be noted that the above-described embodiments can be combined freely as needed. The above merely describes the preferred embodiments of the present application, and it should be noted that, for ordinary skilled in the art, some improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements shall be considered as falling within the protection scope of the present application.

Claims

1. A method for network sparsification measurement based on graph neural network, characterized in that, The method comprises the following steps: S1: sparse sampling of end-to-end network service quality data; S2: based on the sparse sampling data, a network service quality prediction model is established and trained; S3: using a continuous learning function, historical learned spatio-temporal network service quality information is used for future learning and reasoning; In step S2, the network service quality prediction model is established and trained, specifically: S21: the delay matrix of the end-to-end path is constructed into graph structure data; S22: based on the feature encoder of the graph neural network, the features exhibited by the graph structure data are encoded using a multi-layer encoding layer based on message passing; S23: using a delay predictor based on a multi-layer perceptron, the hidden features generated after the graph structure data in step S22 is encoded are restored into a complete network service quality prediction matrix; S24: using a gradient descent method, repeated training optimization is performed to minimize the error; In step S21, the delay matrix of the end-to-end path is constructed into the graph structure data, specifically: The distributed system comprising n network nodes, the graph structure data G(V, E) built, wherein V = {v1, v2, … v n} is a set of all n network nodes, e ij ∈E represents the network service quality from network node i to network node j; In step S22, based on the feature encoder of the graph neural network, the features exhibited by the graph structure data are encoded using a multi-layer encoding layer based on message passing, specifically: For any u∈V, the network service quality information between the neighbor network nodes of u is embedded into the representation of u by aggregation, and the aggregation formula is: where AGG l is an aggregation function, σ is a non-linear function sigmod, CONCAT represents concatenation of two vectors, l is the layer index, v is the node to be encoded, is the learnable parameter for encoding in the l-th layer, is the feature representation of node x in the l-th layer for node u, is the feature representation of the quality of service from node u to node v in the (l-1)-th layer. After the aggregation is completed, the output of the current layer is obtained by combining the information represented by v of the previous layer With the information represented by v of the previous layer, the output of the current layer is obtained The expression is: wherein, is a learnable parameter matrix for the lth layer used for the aggregation, is the output of the l-1th layer, h0 is initialized using the features x of v when the current layer is the first layer; define an inter-layer residual, inject the inter-layer residual into an output y of the first layer l In particular, information of a previous layer is injected into a current layer in a proportion of 1 - λ. wherein, is the output of the previous layer; Define the initial residual error to compensate for gradient vanishing and gradient explosion caused by multi-layer graph convolution coding, and the expression is: Where x0 is the initial feature of all nodes, and β is an adaptive scaling ratio determined using the Frobenius norm; wherein h l is a feature matrix of all nodes of the l-th layer, and x0is an initial feature matrix of all nodes. The inter-layer residual and the initial output are finally superimposed to the output of the graph neural network The output is performed on the above expression: At this time, the encoding of the features exhibited by the graph structure data of the lth layer is completed.

2. The graph neural network based network sparsification measurement method according to claim 1, characterized in that, The end-to-end network service quality data includes end-to-end delay data, end-to-end bandwidth data, and end-to-end packet loss rate data.

3. The graph neural network based network sparsification measurement method according to claim 1, characterized in that, In step S1, the end-to-end network service quality data is sparse sampled, specifically: S11: formulating a network sparsification measurement plan, for a network containing n network nodes, containing n 2 a distributed system of end-to-end paths, in all of the end-to-end paths, randomly sampling according to a sampling rate α, wherein the random sampling is performed in a Bernoulli sampling manner; S12: execute the network sparse measurement plan, and save the sampling results of the network sparse measurement plan to the central node.

4. The graph neural network based network sparsification measurement method according to claim 1, characterized in that, In step S23, using a delay predictor based on a multi-layer perceptron, the hidden features generated after the graph structure data in step S22 is encoded are restored into a complete network service quality prediction matrix, specifically: In step S2, for any one network node v, the implicit feature is generated Using a multi-layer perceptron-based decoding method to reconstruct the service quality matrix, the high-dimensional hidden representation of the service quality matrix is restored, and the expression is: wherein, is the quality of service prediction output for the final model, t is the current time, W is a learnable parameter matrix for the output, T is a transpose operation of the matrix, and b is a bias coefficient matrix.

5. The graph neural network based network sparsification measurement method according to claim 1, characterized in that, In step S24, using a gradient descent method, the error is minimized, specifically: The mean square error is used to describe the difference between the predicted value and the true value, and the expression is: wherein 1 is the error between the predicted value and the true value, A is the predicted quality of service, t A is the true quality of service; Based on the obtained error, the gradient descent method is used for repeated training optimization, and in the training process, the Adam optimizer is used to dynamically adjust the learning rate until the error change between two optimizations is less than a preset threshold.

6. The network sparsification measurement method based on a graph neural network according to claim 1, characterized in that, In step S3, using a continuous learning function, historical learned spatio-temporal network service quality information is used for future learning and reasoning, specifically: F′ n+1 = G(F0, F1,..., F n ) Wherein, G is a continuous learning function, the input is the history model F0, F1…F n , the output is the initial model F' n+1 at the n+1 moment, and the model is initialized by F' n+1 at the n+1 moment, and the training is carried out on the basis of the initialization.

7. A graph neural network-based network sparsification measurement system for performing the graph neural network-based network sparsification measurement method according to any one of claims 1-6, characterized by, It includes: The sparse data collection module is configured to collect sparse data of end-to-end network service quality. The model establishing and training module is configured to establish and train a network service quality prediction model based on the sparse data. The continuous learning module is configured to use historical learned space-time network service quality information for learning and reasoning at future time by using a continuous learning function.

Images