Knowledge fusion graph neural network-based computer network performance evaluation model construction method, system, device and medium

Abstract

The application discloses a computer network performance evaluation model construction method, system, device and medium based on a knowledge fusion graph neural network, and relates to the technical field of computer network performance evaluation model construction.The method comprises the following steps: constructing a knowledge algorithm module, a physical component library, a logical component library, extracting required physical components and logical components, expanding attributes of the physical components and the logical components, encoding the physical components and the logical components after the attributes are expanded, connecting or reconstructing the physical components and the logical components after the attributes are expanded, forming a graph structure, constructing a graph neural network model architecture, training the graph neural network model architecture by using a data set, obtaining a graph neural network model, performing network topology performance index prediction and analysis on the graph neural network model, and obtaining a construction result of the computer network performance evaluation model based on the knowledge fusion graph neural network.The system and the device are used for realizing the computer network performance evaluation model construction method based on the knowledge fusion graph neural network.The application has the characteristics of good interpretability, flexible use site and real reflection of the operation mechanism of a network.

Description

Technical Field

[0001] This invention relates to the field of computer network performance evaluation technology, specifically to a method, system, device, and medium for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network. Background Technology

[0002] As computer networks grow increasingly large, new algorithms for optimizing network performance are constantly emerging. The construction of computer network performance evaluation models plays a crucial role in verifying the feasibility and performance of these optimization algorithms in advance. Building computer network performance evaluation models refers to using various mathematical models or other methods, such as machine learning, to describe, analyze, and optimize the behavior of computer networks. Network performance evaluation models can help us predict network performance, security, and reliability indicators, and provide us with solutions for optimizing network structure and protocols. In the field of computer network performance evaluation, mathematical models were initially used to simulate network behavior, such as queuing systems. However, with the development of machine learning technology, researchers began to apply neural network methods to network modeling. Using neural networks for computer network simulation and modeling is a highly accurate, efficient, and low-cost method. A basic neural network-based computer network simulation model consists of an input layer, hidden layers, and an output layer. Based on this, we can gradually add various protocols and applications to form a more complete simulation model. By embedding computer network knowledge into the network simulation model, we can more realistically simulate the behavior of various devices and protocols in the network, improving the realism and credibility of the simulation model.

[0003] To address the challenge of processing non-Euclidean domain data representing complex relationships between objects, graph neural networks (GNNs) were proposed for handling graph-structured data and have been widely applied in fields such as social networks, knowledge graphs, and transportation systems. GNNs can more accurately and flexibly model the dependencies between nodes in a graph. In computer networks, various entities, such as forwarding nodes and links, have dependencies. Applying GNNs to computer network modeling allows for precise modeling of these dependencies.

[0004] The paper "RouteNet: Leveraging Graph Neural Networks for Network Modeling and Optimization in SDN" abstracts computer networks into a graph structure composed of a series of links and paths. It uses graph neural networks to model the computer network, aggregating messages between links and paths to predict network performance metrics such as latency, jitter, and packet loss rate. The final trained model is then used for latency optimization in SDN scenarios. (In this paper, we propose RouteNet, a novel network model based on Graph Neural Network (GNN) that is able to understand the complex relationship between topology, routing, and input traffic to produce accurate estimates of the per-source / destination per-packet delay distribution and loss. Finally, in order to showcase the potential of our GNN model, we present a series of use cases applicable to an SDN architecture.) Existing neural network performance evaluation models typically only consider basic factors constituting network behavior, such as network topology, traffic, and routing, neglecting the impact of other complex knowledge and algorithms contained within the network. This results in computer network models using neural networks typically having poor interpretability, limited application scenarios, and insufficient flexibility in use. Furthermore, without the support of knowledge and algorithms within computer networks, network models can only be simple combinations of data and algorithms, failing to truly reflect the network's operational mechanism.For example, the paper "RouteNet: Leveraging Graph Neural Networks for Network Modeling and Optimization in SDN" states that (The main assumption behind RouteNet is that information at the path level (e.g., end-to-end metrics such as delays or packet loss) and the link level (e.g., link delay, packet loss rate, link optimization) can be encoded in learnable vectors of real numbers (path and link state vectors respectively). Note that the path abstraction may not necessarily correspond to a physical path. It could be a generic end-to-end traffic flow.) This paper oversimplifies the computer network, only including links and paths. Such a network is conceptually overly generalized and fails to reflect the specificities of the SDN network architecture and its internal scheduling algorithms. Consequently, its latency optimization performance in SDN scenarios is not ideal, and its application is highly limited. It can only be used in the most basic network scenarios and cannot truly reflect the characteristics of the network's operating mechanism.

[0005] Patent application number [202111619382.2] provides a method, apparatus, and device for evaluating the performance of training-free neural networks based on NAS. It obtains the neural network structure by searching based on a NAS-based search strategy, acquires the responses generated by different input samples in all hidden layers during the initialization of the neural network structure, calculates the correlation between the responses, and calculates the network performance score of the neural network structure based on the correlation. However, since the NAS technology it uses is based on the task objective and the corresponding dataset, it directly and automatically generates the deep neural network architecture without the process of manually building the neural network. Therefore, it has the characteristics of poor interpretability and large limitations in application scenarios. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the present invention aims to provide a method, system, device, and medium for constructing a computer network performance evaluation model based on knowledge fusion graph neural networks. This method influences the behavior of individual physical and logical components through knowledge algorithm modules, as well as the connection relationships of graph structures in the graph neural network architecture. By integrating knowledge algorithms into the computer network performance evaluation model constructed based on graph neural networks, the method features good interpretability, flexible application scenarios, and the ability to truly reflect the network's operating mechanism.

[0007] The method for constructing a computer network performance evaluation model based on knowledge fusion graph neural networks includes the following steps:

[0008] Step 1: Based on speech, text, and image in computer network knowledge multimodal, construct a knowledge algorithm module, a physical component library, and a logic component library;

[0009] Step 2: Based on the computer network scenario and the knowledge algorithm module obtained in Step 1, extract the required physical and logical components from the physical and logical component libraries obtained in Step 1.

[0010] Step 3: Expand the attributes of the physical and logical components extracted in Step 2;

[0011] Step 4: Encode the physical and logical components after expanding their attributes in Step 3 to obtain the hidden state vector, as shown in the following formula:

[0012]

[0013] in, Represents the state of a component numbered i at time T, where R represents a router, S represents a switch, T represents a terminal, L represents a link, P represents a path, and F represents a flow.

[0014] Step 5: Connect or reconstruct the physical and logical components after encoding and expanding the attributes in Step 4 to form the graph structure required for the graph neural network input.

[0015] Step 6: Based on the graph structure obtained in Step 5, construct the graph neural network model architecture;

[0016] Step 7: Collect data in the network with 14-24 nodes to obtain the dataset required for training;

[0017] Step 8: Use the dataset obtained in Step 7 to train and evaluate the graph neural network model architecture obtained in Step 6 to obtain an accurate graph neural network model; then input the obtained accurate graph neural network model into a network with 300 or more nodes to perform network topology performance index prediction and analysis, and obtain the construction results of the computer network performance evaluation model based on knowledge fusion graph neural network.

[0018] The knowledge algorithm module in step 1 refers to the description of how the physical and logical components affected by the specific algorithm influence its behavior; the specific algorithm includes transmission control, security measures, management and monitoring methods in the data link layer and network layer.

[0019] The physical components include: node modules, which include routers, switches, and terminals. Their basic characteristics are described below:

[0020] Router: Contains a routing table (N*N matrix), ports, and queues at those ports; represented as:

[0021]

[0022] Among them, R i The characteristics of router i represent the routing table in an N*N topology with N nodes, p j This represents port j in the router. This means that this queue belongs to port j and queue number k;

[0023] Switch: Ports and their connection speeds; represented as:

[0024]

[0025] Among them, S i p represents the characteristics of the switch numbered i. j This represents port j in the switch. This represents the connection link speed of a specific port.

[0026] Terminal: Traffic generation, including traffic characteristics; represented as:

[0027] T i ∈<Markovian;Possion;On-Off;Autocorrelated exponentials...>

[0028] Among them, T i The traffic distribution characteristics of terminal number i;

[0029] Logical components include: links, paths, and data flows, and their basic characteristics are as follows:

[0030] Link: Connected nodes, bandwidth, utilization rate; represented as:

[0031] L i =<{src,dst},c,d / c>

[0032] Among them, L i The link characteristic represented by number i is: src represents the source node, dst represents the destination node, c represents the bandwidth, and d / c represents the utilization rate.

[0033] Path: A series of links representing the forwarding path of a single packet; represented as:

[0034] P i = <l1,l2...,l j >

[0035] Among them, P i Represents the path feature numbered i, l j This indicates that the path contains a link feature with the number j.

[0036] Data flow: Composed of a series of paths; represented as:

[0037] F i = <p1,p2...,p j >

[0038] Among them, F i P represents the flow feature with ID i. j This indicates that the stream contains path feature number j.

[0039] The specific process of expanding the attributes of the extracted physical and logical components in step 3 is as follows:

[0040] Determine the impact of the knowledge algorithm module on the inputs and outputs of the physical and logical components, as well as the behavior of the physical and logical components. Based on this impact, and referring to the following formula, add algorithm-related attributes to the physical and logical components using one-hot codes:

[0041] Algorithm→<Attribute(One-hot code)>

[0042] The aforementioned component behavior impact refers to the different states that components exhibit under the influence of the algorithm, represented by one-hot codes.

[0043] Step 5 connects directly related nodes based on the algorithm knowledge modules and topological connections, abstracting them into the graph structure required for the graph neural network input. Here, the graph structure refers to a data structure composed of nodes and edges. The attributes of the nodes refer to the hidden state vectors of the physical and logical components described in step 4, and the edges refer to the connection and reconstruction processes in steps 5.1-5.3. The specific process of step 5 is as follows:

[0044] Step 5.1: Connect the physical components with the expanded attributes according to the connection method in the topology to obtain the physical local graph structure;

[0045] Step 5.2: Connect the logic components with the expanded attributes to the physical local graph structure obtained in step 5.1 to obtain the physical-logic graph structure;

[0046] Step 5.3: Reconstruct the physical local graph structure obtained in Step 5.1 and the physical-logic graph structure obtained in Step 5.2 according to the knowledge algorithm module to obtain the graph structure required for the graph neural network input; the specific steps of the reconstruction are as follows:

[0047] Step 5.3.1: According to the description in the knowledge algorithm module, establish a connection between the physical local graph structure obtained in step 5.1 and the physical-logic graph structure obtained in step 5.2;

[0048] Step 5.3.2: Delete duplicate and redundant connections from step 5.3.1.

[0049] The specific process of constructing the graph neural network model architecture in step 6 is as follows:

[0050] Step 6.1: Based on the connection relationships between the graph structures required for the graph neural network input obtained in Step 5, determine which component states determine the update of the component state, and initialize the feature vectors of the corresponding physical and logical components, filling them with 0 to make all feature vectors have the same dimension.

[0051] Step 6.2: Aggregate the feature vectors of the relevant components obtained in Step 6.1 through Gated Recurrent Units (GRUs). Each aggregation is considered a state update and is called a message passing.

[0052] Step 6.3: Repeat the message passing process in Step 6.2 N times to form an N-times message passing process, and obtain the higher-order hidden features of the graph.

[0053] Step 6.4: Repeat steps 6.1-6.3 continuously to obtain the fully connected neural network with activation function after the required message is passed, thus obtaining the graph neural network model architecture.

[0054] Step 7 involves data collection in a network with 14-24 nodes to obtain the dataset required for training, specifically as follows:

[0055] In a real network or using a discrete event network simulator, build a small network with 14-24 nodes as needed, collect a dataset, extract a portion of the dataset for training the network model, and use the remainder as a test set to evaluate the generalization error of the model.

[0056] In step 8, the dataset obtained in step 7 is used to train and evaluate the graph neural network model architecture obtained in step 6 to obtain an accurate graph neural network model. The specific process is as follows:

[0057] A portion of the dataset from step 7 is used to train the network model in step 6, and the remaining portion is used as a test set to evaluate the model. During training, the initial learning rate and loss function of the optimizer are set.

[0058] The precise graph neural network model is input into a network with 300 or more nodes to predict and analyze network topology performance indicators, resulting in the construction of a computer network performance evaluation model based on knowledge fusion graph neural networks. The specific process is as follows:

[0059] The model, after training and evaluation, has a prediction accuracy of over 90%. It is used to predict the performance metrics of the target topology to be evaluated, including latency, jitter, and packet loss rate. The predicted performance metrics are analyzed according to the requirements, and appropriate weights are assigned to the performance metrics to evaluate the merits of the algorithm.

[0060] A system for constructing a computer network performance evaluation model based on knowledge fusion graph neural networks includes:

[0061] Physical Component Library, Logical Component Library, and Knowledge Algorithm Module Creation Module: Used to create the physical component library, logical component library, and knowledge algorithm module required for subsequent steps. It transforms computer network knowledge or algorithms into descriptions of the behavioral effects of various physical and logical components in the network, forming the knowledge algorithm module. It uses routers, switches, and terminal physical devices to form the physical component library, and it uses parts of the network that cannot be directly described but need to be abstracted, including links, paths, and data flows, to form the logical component library.

[0062] Graph structure abstraction module: Extract the corresponding physical and logical components according to the knowledge algorithm module and computer network scenario, extend their attributes, encode the hidden state vector, and connect these components according to the knowledge algorithm module and computer network scenario to obtain the graph structure;

[0063] Graph Neural Network Model Building Module: Constructs a graph neural network framework based on the connection relationships of each component in the graph structure, and uses the constructed dataset to train and evaluate the graph neural network to obtain an accurate graph neural network model;

[0064] Network performance evaluation module: Applying accurate graph neural network models to networks with 300 or more nodes to predict and analyze network topology performance indicators, and obtaining the results of building a computer network performance evaluation model based on knowledge fusion graph neural networks.

[0065] A device for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network includes:

[0066] Memory: Used to store the computer program that implements the method for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network;

[0067] A processor is used to implement the aforementioned method for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network when executing the computer program.

[0068] A computer-readable storage medium comprising:

[0069] The computer-readable storage medium stores a computer program that, when executed by a processor, enables a method for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network.

[0070] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0071] 1. This invention introduces the construction of physical component libraries and logical component libraries. It not only organizes commonly used physical devices into a physical component library, but also abstracts and designs some basic but important logical components in the network into a logical component library. Components in the component library can have attributes added according to user needs, so that knowledge algorithms can affect the behavior of individual devices. Compared with the prior art, the simulation of network behavior is more realistic, the application range is wider, and the flexibility of user calls and design usage scenarios is guaranteed.

[0072] 2. Because this invention uses a graph neural network architecture, computer network knowledge can not only affect the behavior of individual devices, but also directly affect the structure of the neural network by influencing the connections of the graph structure. Compared with the prior art, this invention further embeds computer network knowledge into the construction of the computer network model, enhances the interpretability of the computer network model based on the neural network model, and can truly reflect the network's operating mechanism.

[0073] In summary, this invention influences the behavior of individual physical and logical components and the connection relationships of graph structures in graph neural network architecture through a knowledge algorithm module. It integrates knowledge algorithms into a computer network performance evaluation model built on graph neural networks, thus possessing the characteristics of good interpretability, flexible application, and the ability to truly reflect the network's operating mechanism. Attached Figure Description

[0074] Figure 1 This is a flowchart of the present invention.

[0075] Figure 2 This is the overall framework diagram of the present invention.

[0076] Figure 3 This invention provides a transformation graph of the graph structure required to form the input of a graph neural network based on the extracted physical and logical components; wherein, Figure 3 (a) is a data center network scenario. Figure 3 (b) is the physical-logical diagram structure after the data center network scenario is abstracted. Detailed Implementation

[0077] See Figure 1 , Figure 2 A method for constructing a computer network performance evaluation model based on knowledge fusion graph neural networks includes the following steps:

[0078] Step 1: Based on speech, text, and image in computer network knowledge multimodal, construct a knowledge algorithm module, a physical component library, and a logic component library;

[0079] The knowledge algorithm module is formed by transforming computer network knowledge or algorithms into descriptions of the behavioral effects of various physical and logical components in the network; the physical component library is formed by physical devices such as terminals, routers, and switches; and the logical component library is formed by abstracting the parts of the network that cannot be directly described, including links, paths, and data flows.

[0080] The knowledge algorithm module in step 1 refers to the description of how the physical and logical components affected by the specific algorithm influence its behavior; the specific algorithm includes transmission control, security measures, management and monitoring methods in the data link layer and network layer.

[0081] The physical components include: node modules, including routers, switches, and terminals, whose basic characteristics are described below:

[0082] Router: Contains a routing table (N*N matrix), ports, and queues at those ports; represented as:

[0083]

[0084] Among them, R i The characteristics of router i represent the routing table in an N*N topology with N nodes, p j This represents port j in the router. This means that this queue belongs to port j and queue number k;

[0085] Switch: Ports and their connection speeds; represented as:

[0086]

[0087] Among them, S i p represents the characteristics of the switch numbered i. j This represents port j in the switch. This represents the connection link speed of a specific port.

[0088] Terminal: Traffic generation, including traffic characteristics; represented as:

[0089] T i ∈<Markovian;Possion;On-Off;Autocorrelated exponentials...>

[0090] Among them, T i The traffic distribution characteristics of terminal number i;

[0091] Logical components include: links, paths, and data flows, and their basic characteristics are as follows:

[0092] Link: Connected nodes, bandwidth, utilization rate; represented as:

[0093] L i =<{src,dst},c,d / c>

[0094] Among them, L i The link characteristic represented by number i is: src represents the source node, dst represents the destination node, c represents the bandwidth, and d / c represents the utilization rate.

[0095] Path: A series of links representing the forwarding path of a single packet; represented as:

[0096] P i = <l1,l2...,l j >

[0097] Among them, P i Represents the path feature numbered i, l j This indicates that the path contains a link feature with the number j.

[0098] Data flow: Composed of a series of paths; represented as:

[0099] F i = <p1,p2...,p j >

[0100] Among them, F i P represents the flow feature with ID i. j This indicates that the stream contains path feature number j.

[0101] Step 2: Based on the computer network scenario and the knowledge algorithm module obtained in Step 1, extract the required physical and logical components from the physical and logical component libraries obtained in Step 1.

[0102] Step 3: Expand the attributes of the physical and logical components extracted in Step 2; the specific process is as follows:

[0103] Determine the impact of the knowledge algorithm module on the inputs and outputs of the physical and logical components, as well as the behavior of the physical and logical components. Based on this impact, and referring to the following formula, add algorithm-related attributes to the physical and logical components using one-hot codes:

[0104] Algorithm→<Attribute(One-hot code)>

[0105] The aforementioned component behavior impact refers to the different states that components exhibit under the influence of the algorithm, represented by one-hot codes.

[0106] Step 4: Encode the physical and logical components after expanding their attributes in Step 3 to obtain the hidden state vector, as shown in the following formula:

[0107]

[0108] in, Represents the state of a component numbered i at time T, where R represents a router, S represents a switch, T represents a terminal, L represents a link, P represents a path, and F represents a flow.

[0109] The encoding of the physical and logical components after expanding the attributes in step 3 refers to the tensors used in the Tensorflow deep learning library in the Python programming language.

[0110] Step 5 involves connecting or reconstructing the physical and logical components after encoding and expanding their attributes in Step 4 to form the graph structure required for the graph neural network input. Step 5 connects directly related nodes based on algorithm knowledge modules and topological connections, abstracting them into the graph structure required for the graph neural network input. Here, the graph structure refers to a data structure composed of nodes and edges. The attributes of the nodes refer to the hidden state vectors of the physical and logical components mentioned in Step 4, and the edges refer to the connection and reconstruction processes in Steps 5.1-5.3. The specific process of Step 5 is as follows:

[0111] Step 5.1: Connect the physical components with the expanded attributes according to the connection method in the topology to obtain the physical local graph structure;

[0112] Step 5.2: Connect the logic components with the expanded attributes to the physical local graph structure obtained in step 5.1 to obtain the physical-logic graph structure;

[0113] Step 5.3: Reconstruct the physical local graph structure obtained in Step 5.1 and the physical-logic graph structure obtained in Step 5.2 according to the knowledge algorithm module to obtain the graph structure required for the graph neural network input; the specific steps of the reconstruction are as follows:

[0114] Step 5.3.1: According to the description in the knowledge algorithm module, establish a connection between the physical local graph structure obtained in step 5.1 and the physical-logic graph structure obtained in step 5.2;

[0115] Step 5.3.2: Delete duplicate and redundant connections from step 5.3.1.

[0116] Step 6: Based on the graph structure obtained in Step 5, construct the graph neural network model architecture; the specific process is as follows:

[0117] Step 6.1: Based on the connection relationships between the graph structures required to form the graph neural network input obtained in Step 5, determine which component states determine the update of the component states, and initialize the feature vectors of the corresponding physical and logical components, filling them with 0s to make all feature vectors have the same dimension.

[0118] Step 6.2: Aggregate the feature vectors of the relevant components obtained in Step 6.1 through Gated Recurrent Units (GRUs). Each aggregation is considered a state update and is called a message passing.

[0119] Step 6.3: Repeat the message passing process in Step 6.2 N times to form an N-times message passing process, and obtain the higher-order hidden features of the graph.

[0120] Step 6.4: Repeat steps 6.1-6.3 continuously to obtain the fully connected neural network with activation function after the required message is passed, thus obtaining the graph neural network model architecture.

[0121] Step 7: Collect data in the network with nodes 14-24 to obtain the dataset required for training. The specific process is as follows:

[0122] In a real network or using a discrete event network simulator, build a small network with 14-24 nodes as needed, collect a dataset, extract a portion of the dataset for training the network model, and use the remainder as a test set to evaluate the generalization error of the model.

[0123] Step 8: Train and evaluate the graph neural network model architecture obtained in Step 6 using the dataset obtained in Step 7 to obtain an accurate graph neural network model; then input the obtained accurate graph neural network model into a network with 300 or more nodes to predict and analyze network topology performance indicators, and obtain the construction results of the computer network performance evaluation model based on knowledge fusion graph neural network; the specific process is as follows:

[0124] 80% of the dataset from step 7 is used for training the network model in step 6, and the remaining 20% ​​is used as a test set to evaluate the model. During training, the initial learning rate and loss function of the optimizer are set. An Adam optimizer with an initial learning rate of 0.001 is used, and the mean squared error (MSE) is set as the loss function.

[0125] The precise graph neural network model is input into a network with 300 or more nodes to predict and analyze network topology performance indicators, resulting in the construction of a computer network performance evaluation model based on knowledge fusion graph neural networks. The specific process is as follows:

[0126] The model, after training and evaluation, has a prediction accuracy of over 90%. It is used to predict the performance metrics of the target topology to be evaluated, including latency, jitter, and packet loss rate. The predicted performance metrics are analyzed according to the requirements, and appropriate weights are assigned to the performance metrics to evaluate the merits of the algorithm.

[0127] See Figure 3 , Figure 3 (a) represents a data center network scenario, whose network consists of computer terminals, routers, switches, links, flows, and paths. Figure 3 (b) is its abstracted physical-logic diagram structure, where gray lines represent physical connections and blue and red lines represent logical connections.

[0128] A system for constructing a computer network performance evaluation model based on knowledge fusion graph neural networks includes:

[0129] Physical Component Library, Logical Component Library, and Knowledge Algorithm Module Creation Module: Used to create the physical component library, logical component library, and knowledge algorithm module required for subsequent steps. It transforms computer network knowledge or algorithms into descriptions of the behavioral effects of various physical and logical components in the network, forming the knowledge algorithm module. It uses routers, switches, and terminal physical devices to form the physical component library, and it uses parts of the network that cannot be directly described but need to be abstracted, including links, paths, and data flows, to form the logical component library.

[0130] Graph structure abstraction module: Extract the corresponding physical and logical components according to the knowledge algorithm module and computer network scenario, extend their attributes, encode the hidden state vector, and connect these components according to the knowledge algorithm module and computer network scenario to obtain the graph structure;

[0131] Graph Neural Network Model Building Module: Constructs a graph neural network framework based on the connection relationships of each component in the graph structure, and uses the constructed dataset to train and evaluate the graph neural network to obtain an accurate graph neural network model;

[0132] Network performance evaluation module: Applying accurate graph neural network models to networks with 300 or more nodes to predict and analyze network topology performance indicators, and obtaining the results of building a computer network performance evaluation model based on knowledge fusion graph neural networks.

[0133] A device for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network includes:

[0134] Memory: Used to store the computer program that implements the method for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network;

[0135] A processor is used to implement the aforementioned method for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network when executing the computer program.

[0136] A computer-readable storage medium comprising:

[0137] The computer-readable storage medium stores a computer program that, when executed by a processor, enables a method for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network.

[0138] The computer network performance evaluation model based on knowledge fusion graph neural networks is a network model that combines knowledge fusion and graph neural network technology. Its core idea is to influence the behavior of individual network devices and the architecture of the network model through knowledge algorithms, so as to use prior knowledge to help the graph neural network learn the features of network data more accurately, thereby improving the accuracy and efficiency of network performance evaluation.

[0139] The most typical application scenario for computer network performance evaluation models based on knowledge fusion graph neural networks is algorithm evaluation and optimization in communication networks. For example, routing algorithms, flow control algorithms, and other transmission control, security measures, management, and monitoring methods at the data link layer and network layer can be embedded into network models to evaluate their impact on performance (latency, jitter, packet loss rate) in different network scenarios, and then optimize them accordingly.

[0140] Data center networks typically have complex, multi-layered topologies with intricate relationships between nodes. Accurate performance evaluation requires considering various factors, such as network topology, congestion levels, and routing algorithms. Knowledge fusion graph neural network-based computer network performance evaluation models can combine these factors to accurately predict and assess network performance.

[0141] Wireless network performance evaluation needs to consider multiple aspects, such as bandwidth utilization, transmission latency, and energy consumption. A computer network performance evaluation model based on knowledge fusion graph neural networks incorporates these key factors into the construction of algorithm modules and component libraries. This transforms the traditional complex analysis and computation process into a neural network learning process, effectively solving the problems of high complexity and low accuracy associated with traditional methods.

[0142] In the relatively new field of satellite communication networks, computer network performance evaluation models based on knowledge fusion graph neural networks can be applied to many aspects, such as network topology design, link capacity allocation, and scheduling algorithm design. For example, in network topology design, this model can quickly assess the impact of different topologies on network performance based on the network topology and node attributes, and then select the optimal network topology. In link capacity allocation, this model can allocate link capacity in real time based on user demand, channel conditions, and other factors, ensuring network efficiency, stability, and reliability. In scheduling algorithm design, this model can predict network performance based on different scheduling strategies and scenarios, thereby designing more efficient scheduling algorithms.

[0143] The computer network performance evaluation model based on knowledge fusion graph neural networks can also be applied to resource allocation, dynamic routing, and transmission protocol design in satellite communication networks. For example, in resource allocation, the model can optimize resource allocation based on network load and node attributes, improving network resource utilization. In dynamic routing, the model can predict network performance based on different routing strategies and scenarios, thereby selecting the optimal routing path. In transmission protocol design, the model can design transmission protocols more suitable for satellite communication networks based on their characteristics and requirements, improving data transmission efficiency and reliability.

[0144] Furthermore, the construction method of computer network performance evaluation model based on knowledge fusion graph neural networks also shows great potential in some network-like application scenarios. For example, in cloud computing, due to the high dynamism and complexity of cloud computing systems, network performance evaluation has become a critical issue. Traditional network performance evaluation methods are difficult to adapt to the characteristics of cloud computing systems, while the computer network performance evaluation model based on knowledge fusion graph neural networks can specifically address this problem. The computer network performance evaluation model based on knowledge fusion graph neural networks can also be applied to the Internet of Things (IoT) field. With the development of IoT technology, the number of IoT devices is experiencing explosive growth, making IoT network performance evaluation even more important. The computer network performance evaluation model based on knowledge fusion graph neural networks can monitor and analyze IoT devices to achieve performance evaluation and optimization of IoT networks, thereby improving the reliability and stability of IoT systems.

Claims

1. A method for constructing a computer network performance evaluation model based on knowledge fusion graph neural networks, characterized in that, Includes the following steps: Step 1: Based on speech, text, and image in computer network knowledge multimodal, construct a knowledge algorithm module, a physical component library, and a logic component library; Step 2: Based on the computer network scenario and the knowledge algorithm module obtained in Step 1, extract the required physical components and logical components from the physical component library and logical component library obtained in Step 1. Step 3: Expand the attributes of the physical and logical components extracted in Step 2; The specific process is as follows: Determine the impact of the knowledge algorithm module on the inputs and outputs of the physical and logical components, as well as the behavior of the physical and logical components. Based on this impact, and referring to the following formula, add algorithm-related attributes to the physical and logical components using one-hot codes: The aforementioned component behavior impact refers to the different states of a component under the influence of an algorithm, represented by one-hot codes. Step 4: Encode the physical and logical components after expanding their attributes in Step 3 to obtain the hidden state vector, as shown in the following formula: in, Represents a certain number Components in The state at any given moment, Represents a router. Represents a switch. Represents the terminal. Represents the link, Represents the path, Represents flow; Step 5: Connect or reconstruct the physical and logical components after encoding and expanding the attributes in Step 4 to form the graph structure required for the graph neural network input. Step 5 connects directly related nodes based on the algorithm knowledge modules and topological connections, abstracting them into the graph structure required for the graph neural network input. Here, the graph structure refers to a data structure composed of nodes and edges. The attributes of the nodes refer to the hidden state vectors of the physical and logical components described in step 4, and the edges refer to the connection and reconstruction processes in steps 5.1-5.

3. The specific process of step 5 is as follows: Step 5.1: Connect the physical components with the expanded attributes according to the connection method in the topology to obtain the physical local graph structure; Step 5.2: Connect the logic components with the expanded attributes to the physical local graph structure obtained in step 5.1 to obtain the physical-logic graph structure; Step 5.3: Reconstruct the physical local graph structure obtained in Step 5.1 and the physical-logic graph structure obtained in Step 5.2 according to the knowledge algorithm module to obtain the graph structure required for the graph neural network input; the specific steps of the reconstruction are as follows: Step 5.3.1: According to the description in the knowledge algorithm module, establish a connection between the physical local graph structure obtained in step 5.1 and the physical-logic graph structure obtained in step 5.2; Step 5.3.2: Delete duplicate and redundant connections from step 5.3.1; Step 6: Based on the graph structure obtained in Step 5, construct the graph neural network model architecture; Step 7: Collect data in the network with 14-24 nodes to obtain the dataset required for training; Step 8: Use the dataset obtained in Step 7 to train and evaluate the graph neural network model architecture obtained in Step 6 to obtain an accurate graph neural network model; then input the obtained accurate graph neural network model into a network with 300 or more nodes to perform network topology performance index prediction and analysis, and obtain the construction result of the computer network performance evaluation model based on knowledge fusion graph neural network.

2. The method for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network according to claim 1, characterized in that, The knowledge algorithm module in step 1 refers to the description of how the physical and logical components affected by the specific algorithm influence its behavior; the specific algorithm includes transmission control, security measures, management and monitoring methods in the data link layer and network layer. The physical components include: node modules, including routers, switches, and terminals, whose basic characteristics are described below: Router: Contains a routing table, which is an N*N matrix, ports, and queues at those ports; represented as: in, Representative number is Features of routers The routing table in a topology with N nodes. The router is numbered as follows port, This means that this queue belongs to the queue numbered Port, queue number is ; Switch: Ports and their connection speeds; represented as: in, Representative number is The characteristics of the switch, The number of the representative switch is port, This represents the connection link speed of a specific port. Terminal: Traffic generation, including traffic characteristics; represented as: in, The traffic distribution characteristics of terminal number i; Logical components include: links, paths, and data flows, and their basic characteristics are as follows: Link: Connected nodes, bandwidth, utilization rate; represented as: in, Representative number is Link characteristics, Represents the source node. Represents the destination node. Represents bandwidth. Represents occupancy rate; Path: A series of links representing the forwarding path of a single packet; represented as: in, Representative number is Path characteristics, This means that the path contains the number . Link characteristics; Data flow: Composed of a series of paths; represented as: in, Representative number is Flow characteristics, This means that the stream contains the number 1. Path characteristics.

3. The method for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network according to claim 1, characterized in that, The specific process of constructing the graph neural network model architecture in step 6 is as follows: Step 6.1: Based on the connection relationships between the graph structures required for the graph neural network input obtained in Step 5, determine which component states determine the update of the component state, and initialize the feature vectors of the corresponding physical and logical components, filling them with 0 to make all feature vectors have the same dimension. Step 6.2: Aggregate the feature vectors of the relevant components obtained in step 6.1 through the gated recurrent unit (GRU). Each aggregation is taken as a state update step, which is called a message passing step. Step 6.3: Repeat the message passing process in Step 6.2 N times to form an N-times message passing process, and obtain the higher-order hidden features of the graph. Step 6.4: Repeat steps 6.1-6.3 continuously to obtain the fully connected neural network with activation function after the required message is passed, thus obtaining the graph neural network model architecture.

4. The method for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network according to claim 1, characterized in that, Step 7 involves data collection in a network with 14-24 nodes to obtain the dataset required for training, specifically as follows: In a real network or using a discrete event network simulator, build a small network with 14-24 nodes as needed, collect a dataset, extract a portion of the dataset for training the network model, and use the remainder as a test set to evaluate the generalization error of the model.

5. The method for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network according to claim 1, characterized in that, In step 8, the dataset obtained in step 7 is used to train and evaluate the graph neural network model architecture obtained in step 6 to obtain an accurate graph neural network model. The specific process is as follows: A portion of the dataset from step 7 is used to train the network model in step 6, and the remaining portion is used as a test set to evaluate the model. During training, the initial learning rate and loss function of the optimizer are set. The precise graph neural network model is input into a network with 300 or more nodes to predict and analyze network topology performance indicators, resulting in the construction of a computer network performance evaluation model based on knowledge fusion graph neural networks. The specific process is as follows: The model, after training and evaluation, has a prediction accuracy of over 90%. It is used to predict the performance metrics of the target topology to be evaluated, including latency, jitter, and packet loss rate. The predicted performance metrics are analyzed according to the requirements, and appropriate weights are assigned to the performance metrics to evaluate the merits of the algorithm.

6. A computer network performance evaluation model construction system based on knowledge fusion graph neural networks, used to implement the method of claim 1, characterized in that, include: Physical Component Library, Logical Component Library, and Knowledge Algorithm Module Creation Module: Used to create the physical component library, logical component library, and knowledge algorithm module required for subsequent steps. It transforms computer network knowledge or algorithms into descriptions of the behavioral effects of various physical and logical components in the network, forming the knowledge algorithm module. It uses routers, switches, and terminal physical devices to form the physical component library, and it uses parts of the network that cannot be directly described but need to be abstracted, including links, paths, and data flows, to form the logical component library. Graph structure abstraction module: Extract the corresponding physical and logical components according to the knowledge algorithm module and computer network scenario, extend their attributes, encode the hidden state vector, and connect these components according to the knowledge algorithm module and computer network scenario to obtain the graph structure; Graph Neural Network Model Building Module: Constructs a graph neural network framework based on the connection relationships of each component in the graph structure, and uses the constructed dataset to train and evaluate the graph neural network to obtain an accurate graph neural network model; Network performance evaluation module: Applying accurate graph neural network models to networks with 300 or more nodes to predict and analyze network topology performance indicators, and obtaining the results of building a computer network performance evaluation model based on knowledge fusion graph neural networks.

7. A device for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network, characterized in that, include: Memory: for storing a computer program that implements the method for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network as described in any one of claims 1-5; A processor, configured to implement, when executing the computer program, a method for constructing a computer network performance evaluation model based on a knowledge fusion graph neural network as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, include: The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the computer network performance evaluation model construction method based on a knowledge fusion graph neural network as described in any one of claims 1-5.

Images