A network service quality prediction method and device
By combining Pre-Trained Transformer and graph neural network, and utilizing density clustering and sliding window techniques, the spatiotemporal correlation and semantic information of edge cloud servers are constructed, solving the problems of low prediction accuracy and high complexity in existing technologies, and achieving more efficient network service quality prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- PIO CLOUD COMPUTING (SHANGHAI) CO LTD
- Filing Date
- 2023-06-29
- Publication Date
- 2026-08-04
AI Technical Summary
Existing network service quality prediction methods fail to effectively utilize spatiotemporal correlation and semantic information, and are sensitive to abnormal data, resulting in low prediction accuracy. Traditional measurement methods are also highly complex and prone to causing network congestion.
By combining Pre-Trained Transformer and Graph Neural Network, a server distribution map is constructed based on geographical location. Core points are searched using density clustering, a set of linguistic descriptions of node features is generated, semantic embedding vectors are generated using sliding window, and a graph structure is constructed for feature extraction. Finally, a multilayer perceptron is used for prediction.
It improves the accuracy of network service quality prediction, reduces data redundancy and sampling complexity, reduces network congestion, enhances the utilization of spatiotemporal correlation, and comprehensively considers semantic information to improve prediction accuracy.
Smart Images
Figure CN116743605B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of network service technology, and specifically relates to a method and apparatus for predicting network service quality. Background Technology
[0002] Currently, with the emergence of latency-sensitive applications such as virtual reality, augmented reality, and autonomous driving, the demand for distributed systems is increasing. The real-time and high-throughput requirements of these applications are leading service providers to increasingly favor dense and distributed deployments. Against this backdrop, edge cloud, as an emerging edge computing model, has attracted widespread attention. Applications such as virtual reality, augmented reality, and autonomous driving are critical for low-latency responses. Any tiny delay can disrupt the user experience and reduce the real-time performance of the application. Therefore, to meet the requirements of these applications, service providers need to consider how to bring services as close to the user as possible to reduce network transmission latency. Edge cloud computing provides such a solution by deploying computing, storage, and network resources closer to the user at the edge, thereby reducing transmission latency.
[0003] Edge computing provides services closer to users, reducing data transmission distance and time, thus offering faster response times and lower latency. This is crucial for real-time interaction in virtual reality applications, environmental perception in augmented reality applications, and instant decision-making in autonomous driving applications. By pushing computing tasks and data processing to the edge, edge cloud can alleviate the burden on central cloud and, to some extent, address network bottlenecks and bandwidth limitations. As an emerging edge computing model, edge cloud is significant for providing low-latency, high-performance, and high-throughput services. It not only meets the needs of latency-sensitive applications and provides a better user experience but also brings business opportunities and competitive advantages to service providers. Therefore, against the backdrop of increasing demand for distributed systems, edge cloud has become a hot area of research and attention.
[0004] To better optimize resource allocation and scheduling decisions, enable fault diagnosis and rapid response, thereby improving user experience and achieving better network planning, it is essential to predict network service quality (SQS). SQS prediction helps service providers make resource optimization and scheduling decisions. By accurately predicting the SQS of nodes, computing, storage, and network resources can be allocated rationally to meet the real-time and performance requirements of different applications. This improves resource utilization, reduces resource waste, and provides a better user experience. Simultaneously, accurate SQS prediction aids in fault diagnosis and rapid response. When a distributed edge cloud system experiences a failure, predicting node SQS allows for earlier detection and fault location, reducing the impact on users. Accurate SQS prediction enables rapid response and problem resolution, helping to reduce system downtime and improve user satisfaction. SQS prediction also helps optimize user experience. By predicting node SQS, service providers can select the optimal nodes to handle user requests, providing lower latency, higher bandwidth, and more stable connections. This is particularly important for latency-sensitive applications (such as VR, AR, and autonomous driving), enabling a smoother and more realistic experience. By predicting service quality, bottlenecks and performance limitations in the network can be identified, allowing for the optimization of data transmission paths and network topology. This helps improve network reliability, bandwidth utilization, and data transmission efficiency, supporting the overall performance of distributed edge cloud systems.
[0005] Existing network service quality prediction methods can be categorized into two main types: one based on mathematical time series algorithms such as ARIMA and SVR, and the other based on artificial intelligence machine learning algorithms such as RNN and LSTM. Generally speaking, current network service quality prediction methods suffer from the following problems:
[0006] 1. Failure to consider the role of spatiotemporal relationships in service quality prediction. Network service quality data exhibits strong spatiotemporal correlations, particularly the strong correlation between end-to-end latency and physical distance between two servers. Traditional mathematical and artificial intelligence methods cannot effectively utilize these spatiotemporal correlations between nodes, thus affecting the accuracy of network service quality prediction.
[0007] 2. Current algorithms only perform calculations based on the data itself, ignoring the semantics behind the data. Traditional mathematical methods predict future network service quality through autoregression, relying solely on the numerical patterns in the data. Machine learning methods use gradient descent, aiming to reduce errors between data points, but neglect the semantic information of the data itself.
[0008] 3. Mathematical methods have high requirements for data quality and are highly sensitive to outliers; even a small amount of outlier data can significantly impact the overall service quality prediction results. Artificial intelligence-based methods rely heavily on side information from the data. If the data volume is small, AI-based algorithms often fall into overfitting. However, in real-world scenarios, it is difficult to find other side information to assist in predicting network service quality, aside from the data directly being predicted.
[0009] 4. Traditional measurement methods require sampling the entire network, but sampling the entire network is complex and can easily cause network congestion. Summary of the Invention
[0010] To address the above problems, this invention proposes a method and apparatus for predicting network service quality, combining Pre-Trained Transformer and graph neural networks to predict network service quality. The technical solution adopted by this invention to solve the above technical problems is as follows:
[0011] A method for predicting network service quality includes the following steps:
[0012] S1. Construct a server distribution map based on geographical location and use density clustering to search for core points;
[0013] S2, collect QoS data of the edge cloud servers corresponding to the core points as node features, and generate a language description set corresponding to the node features based on the Pre-Trained Transformer model;
[0014] S3 generates feature text sequences for each edge cloud server based on the language description set and node features, and generates corresponding semantic embedding vectors using a sliding window approach.
[0015] S4: Construct a graph structure based on QoS data, and use a graph convolutional neural network to extract features to obtain topological feature vectors;
[0016] S5, the topological feature vector from step S4 and the semantic embedding vector from step S3 are concatenated, and the network service quality of the edge cloud server is predicted using a multilayer perceptron.
[0017] In step S2, before generating the language description set corresponding to the node features based on the Pre-Trained Transformer model, the Pre-Trained Transformer model is initialized using the natural language of the network service quality data.
[0018] In step S3, the expression for the feature text sequence is:
[0019] S=d1f1d2f2....d n f n ;
[0020] In the formula, n represents the characteristic number, d n Representing node features f n The corresponding language description.
[0021] The process of generating the corresponding semantic embedding vector using a sliding window method includes the following steps:
[0022] a. Set the window length and sliding window step size;
[0023] b. Input strings that are the same length as the window, based on the feature text sequence;
[0024] c. Extract the corresponding embedding vector from the window position, and perform average pooling on all embedding vectors to obtain the semantic embedding vector.
[0025] Step S4 includes the following steps:
[0026] S4.1, constructing a graph structure based on QoS between edge cloud servers;
[0027] The graph structure is represented by G(V, E), where V = {v1, v2, ..., v...} i , ..., v m} is the collection of all edge cloud servers, v i Let m represent the i-th edge cloud server, m represent the number of edge cloud servers corresponding to the core point, and e represent the number of edge cloud servers. ij ∈E, representing edge cloud server v i to edge cloud server v j The service quality representation feature vector between them; the service quality representation feature vector is transmitted through the edge cloud server v i to edge cloud server v j The QoS vector between them is obtained by concatenating the corresponding semantic embedding vector;
[0028] S4.2, Construct a graph convolutional neural network including several convolutional networks, wherein the convolutional networks are connected sequentially, and each convolutional network includes a linear mapping layer and a pooling layer;
[0029] S4.3, Input the graph structure from step S4.1 into the graph convolutional neural network to obtain the topological feature vector of each edge cloud server.
[0030] A prediction apparatus based on the above-mentioned network service quality prediction method includes:
[0031] Server filtering module: used to construct a server distribution map based on the geographical location of all edge cloud servers in the edge cloud system, and to search for core points using density clustering methods;
[0032] Language description set construction module: It is used to collect the node features and QoS data of the corresponding edge cloud servers based on the core points obtained by the server filtering module, and generate the language description set corresponding to the node features based on the Pre-Trained Transformer model;
[0033] Semantic Embedding Vector Acquisition Module: This module generates feature text sequences for each edge cloud server based on the language description set and node features obtained from the language description set construction module, and generates corresponding semantic embedding vectors using a sliding window approach.
[0034] Topology feature vector acquisition module: used to construct a graph structure based on QoS data from all edge cloud servers, and use graph convolutional neural networks to extract features to obtain topology feature vectors;
[0035] Service Quality Prediction Module: This module concatenates the topological feature vector obtained from the topological feature vector acquisition module and the semantic embedding vector obtained from the semantic embedding vector acquisition module, and uses a multilayer perceptron to predict the network service quality of the edge cloud server.
[0036] The beneficial effects of this invention are:
[0037] 1. Due to the strong spatiotemporal correlation of service quality information, most of the data in the end-to-end service quality matrix is redundant. This application reduces data redundancy, decreases sampling and prediction complexity, improves measurement speed, and reduces the possibility of network congestion by using geographic location density clustering.
[0038] 2. Fully explore the spatiotemporal correlation of edge cloud network service quality, and use a sliding window approach to model data within T time windows simultaneously in order to capture temporal correlations;
[0039] 3. Using Pre-Trained Transformer as a general computing engine, concrete text containing natural language information is used as auxiliary input, and feature encoding containing semantic information is used as an aid. When predicting network service quality, the semantic information of the data column itself is also included. The semantic information and spatiotemporal information are combined to complete the service quality prediction task, thereby improving the prediction accuracy. Attached Figure Description
[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0041] Figure 1 This is a schematic diagram of the process of the present invention.
[0042] Figure 2 This is a schematic diagram of the density clustering process. Detailed Implementation
[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0044] Graph Neural Networks (GNNs) are a framework that has emerged in recent years, utilizing deep learning to directly learn from graph-structured data. Their superior performance has attracted significant attention and in-depth exploration by researchers. By formulating specific strategies for nodes and edges in a graph, GNNs transform graph-structured data into standardized representations, which are then fed into various neural networks for training, achieving excellent results in tasks such as node classification, edge information propagation, and graph clustering.
[0045] The history of Generative Neural Networks (GNNs) can be traced back to 2005 when Gori et al. first proposed the concept, using Recurrent Neural Networks (RNNs) to process undirected, directed, labeled, and cyclic graphs. Following this, Scarselli et al. and Micheli et al. inherited and developed this model of GNN algorithms, making certain improvements. Early GNNs primarily used RNNs as their main framework, generating vector representations for each node through simple feature mapping and node aggregation, which could not effectively handle the complex and varied graph data in reality. To address this, Bruna et al. proposed applying CNNs to graphs, cleverly transforming the convolution operator to propose the Graph Convolutional Network (GCN), which has spawned many variants. GCNs achieve translation invariance, local awareness, and weight sharing on graphs, providing conceptual guidance and inspiration for the construction and improvement of other GNN frameworks.
[0046] From Gori et al.'s introduction of the GNN concept in 2005, to the emergence of GCN providing an effective processing paradigm for non-Euclidean structured data, and then to the proposal of different GNN framework variants such as GAE, GAT, GRN, and GGN, as well as the application of GNN in various fields, GNN has undergone a process of development from scratch to optimization in both theory and practice. The family of GNN systems has also been continuously evolving and improving. This journey demonstrates the continuous improvement and optimization of GNN algorithms and structures by numerous researchers.
[0047] Pretrained Models (PTMs) learn language knowledge from massive amounts of data and then fine-tune it on a small amount of labeled data, enabling downstream tasks to better learn language features and task-specific knowledge. Early pretraining work focused on transfer learning, transferring knowledge from the original task to the target task. Based on the labeling of the original and target tasks, there are four settings: (both original and target task data are labeled) inductive learning, (original task is labeled, target task is unlabeled) inductive learning, (original task is unlabeled, target task is labeled) self-learning, and (both original and target tasks are unlabeled) unsupervised transfer learning. Self-supervised learning can be considered a variant of unsupervised learning, utilizing the dependency information of the data itself as supervision. It fits well with the NLP domain, which contains a large number of unlabeled sentences, and has therefore become the main setting for pretrained models in the NLP field.
[0048] Language Model (LM): A language model is a probability distribution of a sequence of words. It determines the probability distribution P for a text of length m, assessing the likelihood of the text's existence. Language models are categorized into two types based on the context considered during prediction: autoregressive language models and autoencoder language models. Autoregressive language models predict the next possible word based on the preceding context (left-to-right language model tasks) or the preceding word based on the following context (right-to-left language model tasks). Their advantage is that they can only utilize information from the preceding or following context. Their disadvantage is that they naturally match the goals of some downstream NLP tasks, such as generative tasks, for example: GPT, ELMo (a variant of autoregressive language models that concatenates two preceding and following LSTMs). Autoencoder language models mask random words and predict the word based on its context. Their advantage is that they can naturally integrate contextual information. Their disadvantage is that the model cannot see the masking markers during the fine-tuning stage, thus introducing some error, as seen in BERT. The difference between pre-training and language models: Pre-training is a concept, while a language model is a probabilistic model defined from a task perspective. Language models perform better in pre-training, so they are often chosen as pre-training models.
[0049] Transformer: An encoder-decoder structure employing self-attention, capable of modeling relationships between different words in a sentence. The encoding stage can generate fused representations for all words in parallel, while the decoding stage can only decode one word at a time from left to right, considering previous decoding results at each step. The relationship between pre-training, language models, and Transformer is: idea – task (pre-training objective) – specific model architecture. Because Transformer effectively captures relationships between words, it has gradually become the standard network structure in NLP, often serving as the backbone of PTM (Programmable Module Model). It's important to note that pre-trained language models do not necessarily use the complete Transformer structure; many variations of the Transformer have subsequently emerged. For example, BERT only uses the Transformer encoder structure; GPT only uses the Transformer decoder structure.
[0050] A method for predicting network service quality, such as Figure 1 As shown, it includes the following steps:
[0051] S1. Construct a server distribution map of all edge cloud servers under the edge cloud system based on geographical location, and use density clustering to search for core points;
[0052] Emerging latency-sensitive applications have high requirements for network service quality, and the need for real-time network service quality sampling is also high. The data packets generated by these samplings place significant measurement pressure on the underlying network. However, network service quality characteristics have strong spatiotemporal correlations; for example, if service quality is poor from Beijing to Shanghai, the probability of poor service quality from Tianjin to Shanghai is also high. Therefore, this application uses density clustering to find clusters, using the nodes at the center of each cluster as service quality probe nodes, thus reducing measurement costs.
[0053] Specifically, such as Figure 2As shown, firstly, all servers are scanned. If the number of servers within a radius R is greater than or equal to MinPoints, it is included in the core point list, and its density-reachable points form a corresponding temporary cluster. For each temporary cluster, it is checked whether any of its points are core points. If so, the temporary cluster corresponding to that point is merged with the current temporary cluster to obtain a new temporary cluster. This process is repeated until every point in the current temporary cluster is either not in the core point list or all its density-reachable points are already in that temporary cluster. The temporary cluster is then upgraded to a full cluster. The same merging operation is performed on the remaining temporary clusters until all temporary clusters have been processed, resulting in several clusters. For each cluster, the point most frequently contained in the temporary clusters is selected as the core point of the cluster. The radius R and MinPoints can be dynamically adjusted based on measurement accuracy, which is existing technology and will not be detailed here.
[0054] S2, collect QoS (Quality of Service) data of the edge cloud servers corresponding to the core points as node features, and generate a language description set corresponding to the node features based on the Pre-Trained Transformer model;
[0055] The expression for the node feature is:
[0056] F = {f1, f2, ..., f} n};
[0057] The expression for the language description set is:
[0058] D = {d1, d2, ..., d} n};
[0059] In the formula, n represents the characteristic number, d n Representing node features f n The corresponding language description.
[0060] After obtaining the node features of each node, they are quantized. Data-type features can be directly normalized, while type features can be One-Hot encoded. For example, if the node features are F = {38.5, 112.3, China Unicom}, the corresponding language description set is D = {latitude, longitude, broadband service operator}.
[0061] The QoS data includes bandwidth, latency, latency jitter, latency drift, packet loss rate, availability and / or throughput, where availability refers to the availability status of the server device.
[0062] Preferably, before generating the language description set corresponding to the node features based on the Pre-Trained Transformer model, the Pre-Trained Transformer model is initialized using the natural language of network service quality data. The natural language can be one or more sentences, such as: "This network service quality assessment includes the following aspects: bandwidth, latency, latency jitter, packet loss rate," or a prompt sentence like "This network service quality assessment includes the following aspects: bandwidth, latency, latency jitter, latency drift, packet loss rate, availability, throughput." The purpose of the prompt sentence is to provide the model with a complete contextual environment, ensuring that the model has certain contextual information before performing semantic encoding, thereby improving the quality of semantic embedding.
[0063] S3 generates feature text sequences for each edge cloud server based on the language description set and node features, and generates corresponding semantic embedding vectors using a sliding window approach.
[0064] The expression for the feature text sequence is:
[0065] S=d1f1d2f2....d n f n ;
[0066] Specifically, the window length is first set, and then the stride of the sliding window, i.e., the distance of each slide, is determined based on the window length. Typically, the stride can be set to half the window length to ensure overlap between windows. Based on the feature text sequence, a fixed-length string of length W is input each time, and the corresponding embedding vector is extracted from the window position. The semantic embedding vector is obtained by average pooling all extracted embedding vectors.
[0067] S4. Based on the connection relationships and QoS data between edge cloud servers, a graph structure is constructed. A graph convolutional neural network is used to extract features to obtain a topological feature vector, including the following steps:
[0068] S4.1, constructing a graph structure based on QoS between edge cloud servers;
[0069] The graph structure is represented by G(V, E), where V = {v1, v2, ..., v...} i , ..., v m} is the collection of all edge cloud servers, v i This represents the i-th edge cloud server, where each edge cloud server is considered a node. m represents the number of edge cloud servers corresponding to the core node, and e... ij ∈E, representing edge cloud server v i to edge cloud server v jThe service quality representation feature vector is defined as follows: E represents the set of edges. This service quality representation feature vector is transmitted through an edge cloud server v. i to edge cloud server v j The QoS vector and its corresponding semantic embedding vector are concatenated.
[0070] S4.2, Construct a graph convolutional neural network including several convolutional networks, wherein the convolutional networks are connected sequentially, and each convolutional network includes a linear mapping layer and a pooling layer;
[0071] For any node v i The linear mapping layer aggregates nodes v i The relevant edge attributes yield the feature representation of the node, and the corresponding formula is:
[0072]
[0073] In the formula, Represents node v i The average QoS among all its neighboring nodes u AGG represents the learnable parameter matrix of the linear mapping layer in the l-th convolutional network. l (·) is an aggregate function. Represents node v i The embedding representation in the l-th layer convolutional network, Represents node v i The input features of the l-th layer convolutional network, where u represents node v i The neighboring nodes. When l=1, this input feature is node v. i The semantic embedding vector, when l≠1, is the output of the linear mapping layer of the previous convolutional network.
[0074] The pooling layer is placed after the linear mapping layer. It combines the output of the linear mapping layer with the output of the previous convolutional network layer to obtain the hidden information representation of the nodes. The corresponding formula is:
[0075]
[0076] In the formula, This represents the learnable parameter matrix of the pooling layer in the l-th convolutional network. Represents node v i The implicit information indicates that when l = 1, That is, node v i The semantic embedding vector, when l≠1, This is the output of the (l-1)th convolutional network layer, which is also the output of the pooling layer in the (l-1)th convolutional network layer.
[0077] S4.3, Input the graph structure from step S4.1 into the graph convolutional neural network to obtain the topological feature vector of each node;
[0078] The convolutional networks are connected sequentially, with the output of the previous convolutional network serving as the input of the next convolutional network. This process, through a graph convolutional neural network, yields a feature vector in the form of a topological information encoding for each node.
[0079] S5, the topological feature vector from step S4 and the semantic embedding vector from step S3 are concatenated, and a multilayer perceptron is used to predict the network service quality of the edge cloud server. The expression is as follows:
[0080]
[0081] This application performs comprehensive prediction based on vectors embedded by semantic embedding and graph neural network embedding, and outputs the corresponding QoS data.
[0082] This application also provides a network service quality prediction device, including:
[0083] Server filtering module: used to construct a server distribution map based on the geographical location of all edge cloud servers in the edge cloud system, and to search for core points using density clustering methods;
[0084] Language description set construction module: It is used to collect the node features and QoS data of the corresponding edge cloud servers based on the core points obtained by the server filtering module, and generate the language description set corresponding to the node features based on the Pre-Trained Transformer model;
[0085] Semantic Embedding Vector Acquisition Module: This module generates feature text sequences for each edge cloud server based on the language description set and node features obtained from the language description set construction module, and generates corresponding semantic embedding vectors using a sliding window approach.
[0086] Topology feature vector acquisition module: used to construct a graph structure based on QoS data from all edge cloud servers, and use graph convolutional neural networks to extract features to obtain topology feature vectors;
[0087] Service Quality Prediction Module: This module concatenates the topological feature vector obtained from the topological feature vector acquisition module and the semantic embedding vector obtained from the semantic embedding vector acquisition module, and uses a multilayer perceptron to predict the network service quality of the edge cloud server.
[0088] This application also provides an electronic device, which includes a processor and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it implements the network service quality prediction method as described above.
[0089] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the network service quality prediction method described above. Specifically, the storage medium can be a general-purpose storage medium, such as a removable disk or hard disk, and the computer program on the storage medium, when run, can execute embodiments of the network service quality prediction method described above.
[0090] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting network service quality, characterized in that, Includes the following steps: S1. Construct a server distribution map based on geographical location and use density clustering to search for core points; S2, collect QoS data of the edge cloud servers corresponding to the core points as node features, and generate a language description set corresponding to the node features based on the Pre-TrainedTransformer model; S3 generates feature text sequences for each edge cloud server based on the language description set and node features, and generates corresponding semantic embedding vectors using a sliding window approach. S4: Construct a graph structure based on QoS data, and use a graph convolutional neural network to extract features to obtain topological feature vectors; S5, the topological feature vector in step S4 and the semantic embedding vector in step S3 are concatenated, and the network service quality of the edge cloud server is predicted by the multilayer perceptron. Step S4 includes the following steps: S4.1, constructing a graph structure based on QoS between edge cloud servers; The graph structure adopts express, It is a collection of all edge cloud servers. Indicates the first Each edge cloud server acts as a node. This indicates the number of edge cloud servers corresponding to the core node. Represents the set of edges; S4.2, Construct a graph convolutional neural network including several convolutional networks, wherein the convolutional networks are connected sequentially, and each convolutional network includes a linear mapping layer and a pooling layer; For any edge cloud server The linear mapping layer aggregates data with edge cloud servers. The relevant edge attributes yield the feature representation of the edge cloud server, and the corresponding formula is: ； In the formula, Indicates edge cloud server With all its neighboring edge cloud servers The average QoS between Indicates the first The learnable parameter matrix of the linear mapping layer in a convolutional network. For aggregate functions, Indicates edge cloud server In the Embedded representations of multilayer convolutional networks Indicates edge cloud server In the Input features of convolutional networks Indicates edge cloud server The neighboring edge cloud server; The pooling layer is placed after the linear mapping layer. It combines the output of the linear mapping layer with the output of the previous convolutional network layer to obtain the hidden information representation of the nodes. The corresponding formula is: ； In the formula, Indicates the first The learnable parameter matrix of pooling layers in a multilayer convolutional network. Indicates edge cloud server Implicit information representation; S4.3, Input the graph structure from step S4.1 into the graph convolutional neural network to obtain the topological feature vector of each edge cloud server.
2. The network service quality prediction method according to claim 1, characterized in that, Before generating the language description set corresponding to node features based on the Pre-TrainedTransformer model, the Pre-TrainedTransformer model is initialized using the natural language of network service quality data.
3. The network service quality prediction method according to claim 1, characterized in that, In step S3, the expression for the feature text sequence is: ； In the formula, Represents the characteristic number, Representing node features The corresponding language description.
4. The network service quality prediction method according to claim 1, characterized in that, The process of generating the corresponding semantic embedding vector using a sliding window method includes the following steps: a. Set the window length and sliding window step size; b. Input strings that are the same length as the window, based on the feature text sequence; c. Extract the corresponding embedding vector from the window position, and perform average pooling on all embedding vectors to obtain the semantic embedding vector.
5. A prediction apparatus based on the network service quality prediction method according to any one of claims 1-4, characterized in that, include: Server filtering module: used to construct a server distribution map based on the geographical location of all edge cloud servers in the edge cloud system, and to search for core points using density clustering methods; Language description set construction module: It is used to collect the node features and QoS data of the corresponding edge cloud servers based on the core points obtained by the server filtering module, and generate the language description set corresponding to the node features based on the Pre-Trained Transformer model; Semantic Embedding Vector Acquisition Module: This module generates feature text sequences for each edge cloud server based on the language description set and node features obtained from the language description set construction module, and generates corresponding semantic embedding vectors using a sliding window approach. Topology feature vector acquisition module: used to construct a graph structure based on QoS data from all edge cloud servers, and use graph convolutional neural networks to extract features to obtain topology feature vectors; Service Quality Prediction Module: This module concatenates the topological feature vector obtained from the topological feature vector acquisition module and the semantic embedding vector obtained from the semantic embedding vector acquisition module, and uses a multilayer perceptron to predict the network service quality of the edge cloud server.