Service network function node identification method based on subgraph information bottleneck

Abstract

The invention provides a service network function node identification method based on subgraph information bottleneck, and belongs to the technical field of networks. The method comprises the following steps: acquiring a topological graph of a to-be-tested service network, and acquiring a feature matrix and an adjacent matrix of the topological graph; the feature matrix and the adjacency matrix are input into a feature coding model, node features of the service network to be tested are obtained through the feature coding model according to the feature matrix and the adjacency matrix, and the feature coding model is obtained after training of an optimization function combining multi-order subgraph information bottleneck and a global topology constraint function; and obtaining a node identification result of the to-be-tested service network according to the node characteristics. Therefore, network node features with multi-scale constraint information in a hidden space are obtained in a data-driven form by combining the multi-level node semantic information and the multi-scale topological information of the sub-graph from the perspective of an information theory, and more accurate functional node features are captured, so that the accuracy of service network node identification is greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of network technology, and in particular to a method for identifying service network function nodes based on subgraph information bottlenecks. Background Art

[0002] A service resource network is a comprehensive computer network system that integrates information service systems, network management systems, and resource information users. By integrating different types of resources, the system achieves efficient management and dynamic allocation of resources, thereby providing comprehensive information services to various resource information users, thereby improving the performance of aspects such as data storage, network services, and system management. The service resource network allocates and schedules resources based on real-time demand, significantly improving resource utilization. Intelligent service resource networks can also monitor and analyze resource usage in real time, predict future demand, and help managers make resource plans in advance.

[0003] In a service resource network, a typical service node is usually an integrated, multifunctional service node. For such multifunctional service nodes, it is necessary to perform functional identification of the service resource network to obtain its functional labels in different application scenarios and prepare data identification for subsequent downstream tasks. In a service resource network, the task of node identification is to identify and classify key nodes with different functions by analyzing the characteristics of each node in the network and the connection relationship between them. This process usually requires comprehensive consideration of the topological properties of the nodes (such as degree, path length, etc.), functional properties (such as bandwidth, latency, etc.), and the interaction relationship between nodes. Through node identification, researchers can better understand the structure of the network, optimize resource allocation, and improve the security and robustness of the network.

[0004] Common related technologies include: (1) node identification methods based on feature engineering and classification algorithms; (2) node identification methods based on deep learning or graph neural networks. However, these methods are difficult to accurately capture the characteristics of functional nodes in service resource networks, resulting in low identification accuracy. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a service network functional node identification method based on subgraph information bottleneck, which combines the multi-level node semantic information and multi-scale topological information of the subgraph in a data-driven form from the perspective of information theory to obtain network node characteristics with multi-scale constraint information in the latent space, and can capture more accurate functional node characteristics in the service resource network, thereby improving the accuracy of service network node identification.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a method for identifying a service network function node based on a subgraph information bottleneck, the method comprising: Obtaining a topological map of the service network to be tested, and obtaining a characteristic matrix and an adjacency matrix of the topological map; wherein the topological map includes a plurality of functional nodes; Inputting the feature matrix and the adjacency matrix into a feature coding model, and using the feature coding model to obtain node features of the service network to be tested according to the feature matrix and the adjacency matrix; wherein the feature coding model is a model obtained by training the optimization function of the joint multi-order subgraph information bottleneck and the global topology constraint function; A node identification result of the service network to be tested is obtained according to the node characteristics.

[0007] Optionally, the step of obtaining the feature encoding model includes: Obtaining matrix information and real features of the sample service network, as well as real connection relationships between functional nodes in the sample service network; wherein the matrix information includes an adjacency matrix and a feature matrix; Using the real features and the real connection relationship as labels of the matrix information; The matrix information is used as the input of the target network, and iterative training is performed using a joint function to obtain a feature encoding model; wherein the target network includes a feature encoding network for learning the features of each functional node and an attention network for learning the connection relationship between functional nodes; the joint function includes an optimization function of a multi-order subgraph information bottleneck and a global topology constraint function, the global topology constraint function is used to calculate the loss value between the real connection relationship and the output result of the attention network, and the optimization function is used to calculate the loss value between the real feature and the output result of the feature encoding network.

[0008] Optionally, the step of using the matrix information as the input of the target network and performing iterative training using a joint function to obtain a feature encoding model includes: Inputting the matrix information into the feature encoding network and the attention network respectively, obtaining a first prediction result output by the target network and a second prediction result output by the attention network; Based on the real connection relationship and the second prediction result, adopting the global topology constraint function, obtaining a global topology loss value; Based on the real feature and the first prediction result, using the optimization function to obtain an information bottleneck optimization value; Combining the global topology loss value and the information bottleneck optimization value to obtain a joint loss value; According to the joint loss value and back propagation, updating the parameters of the feature encoding network and the attention network; When the joint loss value converges, using the current feature encoding network as a feature encoding model; When the joint loss value does not converge, return to the step of inputting the matrix information into the feature encoding network and the attention network respectively to obtain a first prediction result output by the target network and a second prediction result output by the attention network.

[0009] Optionally, the optimization function includes a variable cross entropy function and a KL divergence function, the real feature includes a real adjacency matrix of a first-order subgraph and a second-order subgraph of each functional node, and the first prediction result includes a learned adjacency matrix of a first-order subgraph and a second-order subgraph of each functional node; The step of obtaining the information bottleneck optimization value by using the optimization function based on the real feature and the first prediction result comprises: According to the learned adjacency matrix and the real adjacency matrix of the first-order subgraph of each functional node, the KL divergence function is used to obtain a first-order distribution difference value; According to the learned adjacency matrix and the real adjacency matrix of the second-order subgraph of each functional node, the variable cross entropy function is used to obtain a second-order distribution difference value; The information bottleneck optimization value is obtained by combining the first-order distribution difference value and the second-order distribution difference value.

[0010] Optionally, the optimization function includes: ; in, represents the total number of functional nodes of the sample service network, Characterization The learned adjacency matrix of the second-order subgraph of function nodes, Characterization The true adjacency matrix of the second-order subgraph of function nodes, represents the difference value of the second-order distribution, represents the hyperparameters of the feature encoding network, Characterization The learned adjacency matrix of the first-order subgraph of function nodes, Characterization The true adjacency matrix of the first-order subgraph of function nodes, Represents the first-order distribution difference value.

[0011] Optionally, the second prediction result includes a low-dimensional node representation of each functional node; The step of obtaining a global topology loss value based on the real connection relationship and the second prediction result by using the global topology constraint function comprises: Obtaining a learning connection relationship between each functional node in the sample service network according to each low-dimensional node representation; Obtaining a connection difference value according to the learned connection relationship and the real connection relationship; Based on the low-dimensional node representations and graph variational inference, obtaining a connection distribution difference value; A global topology loss value is obtained according to the connection difference value and the connection distribution difference value.

[0012] Optionally, the step of obtaining a connection distribution difference value based on each of the low-dimensional node representations and graph variational inference includes: Calculating the mean and standard deviation of each of the low-dimensional node representations, and obtaining the variational distribution of each of the low-dimensional node representations according to the mean and standard deviation; Obtaining forward propagation results according to the low-dimensional node representations, the feature matrix and the multi-layer perception; A connection distribution difference value is obtained according to the variational distribution, the forward propagation result and the KL divergence function.

[0013] Optionally, the step of obtaining a learning connection relationship between each functional node in the sample service network based on each low-dimensional node representation includes: Any two of the low-dimensional node representations are input into the interior point function to obtain the learning connection relationship between the functional nodes corresponding to the two low-dimensional nodes.

[0014] Optionally, the global topology constraint function includes: ; Among them, Represents a function node With function node The real connection between , Indicates A low-dimensional node representation, Characterize the interior point function, Represents a function node With function node The learning connection between , represents the total number of low-dimensional node representations, Characterize the feature matrix, Characterize the forward propagation model, and Represent the mean and standard deviation of the low-dimensional node representation respectively.

[0015] Optionally, the step of obtaining a node identification result of the service network to be tested according to the node characteristics includes: The node features are input into a support vector machine to obtain classification results of each functional node in the service network to be tested.

[0016] In a second aspect, the present invention provides a service network function node identification device based on subgraph information bottleneck, including a network modeling module, a feature extraction module and a node identification module; The network modeling module is used to obtain a topology map of the service network to be tested, and obtain a feature matrix and an adjacency matrix of the topology map; wherein the topology map includes a plurality of functional nodes; The feature extraction module is used to input the feature matrix and the adjacency matrix into a feature coding model, and use the feature coding model to obtain the node features of the service network to be tested according to the feature matrix and the adjacency matrix; wherein the feature coding model is a model obtained by training the optimization function of the joint multi-order subgraph information bottleneck and the global topology constraint function; The node identification module is used to obtain the node identification result of the service network to be tested according to the node characteristics.

[0017] In a third aspect, the present invention provides an electronic device comprising a processor and a memory, wherein the memory stores a computer program that can be executed by the processor, and the processor can execute the computer program to implement the service network function node identification method based on subgraph information bottleneck as described in the first aspect.

[0018] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the service network function node identification method based on subgraph information bottleneck as described in the first aspect.

[0019] The method for identifying functional nodes of a service network based on a subgraph information bottleneck provided by an embodiment of the present invention comprises: obtaining a topological map of a service network to be tested, and obtaining a feature matrix and an adjacency matrix of the topological map; wherein the topological map includes a plurality of functional nodes; inputting the feature matrix and the adjacency matrix into a feature coding model, and using the feature coding model to obtain node features of the service network to be tested according to the feature matrix and the adjacency matrix, wherein the feature coding model is a model obtained after training with an optimization function of a joint multi-order subgraph information bottleneck and a global topological constraint function; and obtaining node identification results of the service network to be tested according to the node features. In this way, in a data-driven form, from the perspective of information theory, combining the multi-level node semantic information of the subgraph and the multi-scale topological information, the network node features with multi-scale constraint information in the latent space are obtained, which can capture more accurate functional node features in the service network, thereby greatly improving the accuracy of service network node identification.

[0020] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0022] Figure 1 A schematic diagram of the system architecture of a service network function node identification system provided by an embodiment of the present invention is shown.

[0023] Figure 2 A schematic diagram of a module architecture of an electronic device provided by an embodiment of the present invention is shown.

[0024] Figure 3 One of the flow diagrams of the service network function node identification method provided by an embodiment of the present invention is shown.

[0025] Figure 4 The second flowchart of the service network function node identification method provided by the embodiment of the present invention is shown.

[0026] Figure 5 Shows Figure 4 Schematic diagram of the process of some sub-steps of step 25.

[0027] Figure 6 Shows Figure 5 A flowchart of some sub-steps of step 252 in FIG.

[0028] Figure 7 Shows Figure 5 A flowchart of some sub-steps of step 253.

[0029] Figure 8 Shows Figure 7 A flowchart of some sub-steps of step 253-5.

[0030] Fig. 9 A block diagram of a service network function node identification device provided by an embodiment of the present invention is shown.

[0031] Explanation of the accompanying drawings: 10 - service network function node identification system; 110 - identification device; 120 - model training device; 20 - electronic device; 210 - memory; 220 - processor; 230 - communication module; 30 - service network function node identification device; 310 - network modeling module; 320 - feature extraction module; 330 - node identification module. DETAILED DESCRIPTION

[0032] In different fields, service networks are a type of data with obvious non-Euclidean characteristics. In addition to the semantic characteristics of the network, the topological characteristics of the network are also aspects that need to be considered. However, in the actual process of service network function identification, it is still rare to take into account both the attribute characteristics and topological characteristics of the network to work together to complete the task of service network node identification. Only relying on the single attribute characteristics of the node to obtain node representation, or only relying on the statistical characteristics of the topological structure to analyze the network, which still faces the problem of low efficiency of large-scale network calculation and difficulty in capturing node characteristics in dynamic network environments.

[0033] In order to solve the above problems, an embodiment of the present invention provides a service network functional node identification method based on subgraph information bottleneck. It combines the multi-level node semantic information and multi-scale topological information of the subgraph in a data-driven form from the perspective of information theory to obtain network node characteristics with multi-scale constraint information in the latent space. It can capture more accurate functional node characteristics in the service resource network and improve the accuracy of service network node identification.

[0034] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0035] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0036] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0037] The service network function node identification method based on subgraph information bottleneck provided by the embodiment of the present invention can be applied to Figure 1 In the service network function node identification system 10 shown, the service network function node identification system 10 includes an identification device 110 and a model training device 120. The identification device 110 can be communicated with the model training device 120 via a network, a data cable or other wired or wireless method.

[0038] The model training device 120 is used to train and obtain a feature encoding model, and deploy the feature encoding model to the recognition device 110.

[0039] The identification device 110 is used to implement the service network function node identification method based on the subgraph information bottleneck provided by an embodiment of the present invention based on the feature coding model, namely: obtaining a topological map of the service network to be tested, and obtaining a feature matrix and an adjacency matrix of the topological map; wherein the topological map includes multiple functional nodes; inputting the feature matrix and the adjacency matrix into the feature coding model, and using the feature coding model to obtain the node characteristics of the service network to be tested according to the feature matrix and the adjacency matrix; wherein the feature coding model is a model obtained after training with the optimization function of the joint multi-order subgraph information bottleneck and the global topology constraint function; according to the node characteristics, obtaining the node identification result of the service network to be tested.

[0040] The recognition device 110 may be, but is not limited to: a personal computer, a laptop, a tablet computer, an independent server, a server cluster, a mobile terminal, a wearable portable device, etc. The model training device 120 may be, but is not limited to: an independent server, a server cluster, a personal computer, etc.

[0041] For example, the recognition device 110 and the model training device 120 may be the same device, that is, the model training and the node recognition are performed on the same device. The recognition device 110 and the model training device 120 may also be independent devices.

[0042] Please refer to Figure 2 , is a block diagram of an electronic device 20, which may be Figure 1 The identification device 110 in the service network function node identification system 10 shown may also be a model training device 120. The electronic device 20 includes a memory 210, a processor 220, and a communication module 230. The memory 210, the processor 220, and the communication module 230 are electrically connected to each other directly or indirectly to achieve data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0043] The memory 210 is used to store programs or data and can be, but not limited to, a random access memory, a read-only memory, a programmable read-only memory, an erasable read-only memory, an electrically erasable read-only memory, and the like.

[0044] The processor 220 is used to read / write data or programs stored in the memory 210 and execute corresponding functions. For example, Figure 1 In the service network function node identification system 10 shown, the processor 220 of the identification device 110 executes the computer program stored in the memory 210 to implement the service network function node identification method based on subgraph information bottleneck provided in an embodiment of the present invention, and the processor 220 of the model training device 120 executes the computer program stored in the memory 210 to obtain a feature coding model.

[0045] The communication module 230 is used to establish a communication connection between the electronic device 20 and other communication terminals through the network, and to send and receive data through the network. Figure 1 In the service network function node identification system 10 shown, the communication module 230 of the identification device 110 or the model training device 120 is used to send and receive data through the network.

[0046] It should be understood that Figure 2 The structure shown is only a schematic diagram of the structure of the electronic device 20. The electronic device 20 may also include Figure 2 More or fewer components as shown, or with Figure 2 Different configurations are shown. Figure 2 Each component shown in the figure can be implemented by hardware, software or a combination thereof.

[0047] Reference Figure 3 , a method for identifying service network function nodes based on subgraph information bottleneck provided by an embodiment of the present invention includes steps 11 to 15. Thus,: Figure 1 In the service network function node identification system 10 shown in FIG. Figure 2The structure shown implements steps 11 to 15 when the processor 220 executes the computer program stored in the storage.

[0048] Step 11: Obtain a topological map of the service network to be tested, and obtain a characteristic matrix and an adjacency matrix of the topological map.

[0049] Among them, the topology diagram includes multiple functional nodes, that is, it is composed of the connection relationship between functional nodes. A functional node is an integrated, multi-functional service node in the service network. For example, it can be a client, web server, application server, data server, switch, router, firewall, DNS server, etc.

[0050] Step 13: Input the feature matrix and the adjacency matrix into the feature coding model, and use the feature coding model to obtain the node features of the service network to be tested according to the feature matrix and the adjacency matrix.

[0051] Among them, the feature encoding model is a model obtained after training the optimization function of the joint multi-order subgraph information bottleneck and the global topological constraint function. That is, in the process of calculation and processing, the feature encoding model can combine the multi-level node semantic information and multi-scale topological information of the subgraph from the perspective of information theory to obtain functional node features with multi-scale constraint information in the latent space.

[0052] Step 15: Obtain node identification results of the service network to be tested according to the node characteristics.

[0053] The feature matrix includes the topological graph (which can be expressed as ) contains the preliminary characteristics of each functional node in the network, such as function (such as computing, storage, retrieval, etc.), topological structure, connection distance, running status, service resources, IP address, etc. The adjacency matrix includes the adjacency relationship of each functional node, that is, it is an n×n matrix, in which each element Represents a function node and function nodes For example, if the function node and function nodes There is an edge between them, then The value is 1, if the function node and function nodes There is no edge between them, then The value of is 0.

[0054] For example, in combination Figure 1In the service network function node identification system 10 shown, the model training device 120 trains the target network (such as a graph convolutional neural network) by combining the optimization function of the multi-order subgraph information bottleneck and the global topology constraint function, obtains a feature coding model, and downloads and deploys the feature coding model to the identification device 110. When it is necessary to identify the service network to be tested, the identification device 110 models the service network to be tested, obtains a topology map, and uses a preset feature extraction rule or model to obtain a preliminary feature matrix and adjacency matrix of the topology map. Thus, the identification device 110 inputs the feature matrix and the adjacency matrix into the feature coding model, and uses the feature coding model to obtain the node features of the service network to be tested based on the feature matrix and the adjacency matrix, combining the multi-level node semantic information of the subgraph (i.e., the feature matrix) and the multi-scale topology information (i.e., the adjacency matrix) from the perspective of information theory.

[0055] At this time, the node characteristics include the node characteristics of each functional node in the service network to be tested. Further, the identification device 110 obtains the node identification result of the service network to be tested, that is, the identification result of each functional node, according to the node characteristics.

[0056] Compared with the related technologies that only rely on single node attribute features to obtain node representation, or only rely on topological structure statistical features to analyze the network, the above-mentioned service network functional node identification method based on subgraph information bottleneck adopts a data-driven form and combines the subgraph multi-level node semantic information and multi-scale topological information from the perspective of information theory to obtain network node characteristics with multi-scale constraint information in latent space, which can capture more accurate functional node characteristics in the service network, thereby greatly improving the accuracy of service network node identification.

[0057] The method of training to obtain the feature encoding model can be flexibly selected. For example, a joint function consisting of a multi-order subgraph information bottleneck optimization function and a global topology constraint function can be used to iteratively train the graph convolutional neural network, or training can be performed according to preset rules. The above methods are only examples, and their implementation methods are not limited.

[0058] In order to enable the feature encoding model to learn: from the perspective of information theory, the multi-level node semantic information and multi-scale topological information of the subgraph are combined to obtain the functional node features with multi-scale constraint information in the latent space. In the model training process, the joint function is introduced, and the attention network is used for auxiliary reinforcement learning to train the conception of the feature encoding model. Figure 4 , the process of obtaining the feature encoding model includes steps 21 to 25. Thus,: Figure 1 In the service network function node identification system 10 shown in FIG. Figure 2The structure shown implements steps 21 to 25 when the processor 220 executes the computer program stored in the storage.

[0059] Step 21, obtaining matrix information and real features of the sample service network, and real connection relationships between functional nodes in the sample service network.

[0060] The matrix information includes the adjacency matrix and feature matrix of the topological graph of the sample service network, and the topological graph includes each functional node in the sample service network and the connection relationship between the functional nodes in at least one application scenario. The real connection relationship includes the connection relationship displayed in the topological graph and the implicit connection relationship that is not displayed in the topological graph, that is, the connection relationship between each functional node in various application scenarios (such as storage scenarios, instant messaging, remote login services, data query services, content distribution services, etc.). The real features include the features of each functional node in different scenarios, such as functions (such as computing, storage, retrieval, etc.), topological structure, connection distance, operating status, service resources, IP address, etc.

[0061] Step 23, using the real features and the real connection relationship as labels of the matrix information.

[0062] Step 25, using the matrix information as the input of the target network, using the joint function for iterative training to obtain a feature encoding model.

[0063] The target network includes a feature encoding network for learning the features of each functional node and an attention network for learning the connection relationship between functional nodes. The feature encoding network can be any graph convolutional neural network, and the attention network can be any graph attention network.

[0064] The joint function includes the optimization function of the multi-order subgraph information bottleneck and the global topology constraint function. Among them, the global topology constraint function is to calculate the loss value between the real connection relationship and the output result of the attention network. The output result of the attention network includes the learned connection relationship (that is, the connection relationship between each functional node in the learned sample service network). The optimization function is to calculate the loss value between the real feature and the output result of the feature encoding network. The output result of the feature encoding network includes the features of each functional node in the learned sample service network in each application scenario.

[0065] During the iterative training process, the gradient of the model parameters is calculated based on the loss value between the real features and the output results of the feature encoding network, as well as the loss value between the real connection relationship and the output results of the attention network. Then, the model parameters of the feature encoding network (i.e., graph convolutional neural network) and the attention network are continuously adjusted according to the gradient and learning rate, such as the weights of each layer, optimizer parameters, regularization parameters, dropout rate, etc.

[0066] In the above steps 21 to 25, the attention network is introduced to assist the feature encoding network in enhanced learning, and the feature encoding network is continuously optimized according to the loss value between the real feature and the output result of the feature encoding network, and the loss value between the real connection relationship and the output result of the attention network, so that the feature encoding network (i.e., the feature encoding model) can quickly learn: from the perspective of information theory, combining the multi-level node semantic information and multi-scale topological information of the subgraph, to obtain the functional node features with multi-scale constraint information in the latent space. Therefore, the feature encoding model can mine the node features in the latent space, making the node features more accurate.

[0067] The iterative training method in step 25 can be flexibly set, for example, the number of iterations can be set, and the training is completed when the number of iterations reaches the set value, or the training is terminated when the joint function converges. The above method is only an example, and its implementation method is not limited.

[0068] In order to ensure that the feature encoding model with the required performance (such as accuracy) is obtained, the training is ended when the joint function converges in the iterative training of step 25 to obtain the concept of the feature encoding model. Figure 5 , the iterative training process in step 25 includes steps 251 to 257.

[0069] Step 251, input the matrix information into the feature encoding network and the attention network respectively to obtain the first prediction result output by the target network and the second prediction result output by the attention network.

[0070] Step 252: Based on the real features and the first prediction result, an optimization function is used to obtain an information bottleneck optimization value.

[0071] Step 253: Based on the real connection relationship and the second prediction result, a global topology constraint function is used to obtain a global topology loss value.

[0072] Step 254, combining the global topology loss value and the information bottleneck optimization value to obtain a joint loss value.

[0073] Step 255, update the parameters of the feature encoding network and the attention network according to the joint loss value and back propagation.

[0074] Step 256, determine whether the joint loss value converges. If yes, execute step 257, if no, return to execute step 251.

[0075] Step 257, using the current feature encoding network as a feature encoding model.

[0076] In order to enable the final feature encoding model to mine the node features in the latent space, during initialization, the component needs to initialize the network parameters of the feature encoding model and initialize the first-order subgraph and multi-order subgraph of each functional node according to the adjacency matrix in the matrix information, for example, including at least a second-order subgraph.

[0077] The first-order subgraph of a functional node includes the node itself (i.e., vertex), the adjacent nodes of the vertex, and the edges between the vertex and the adjacent nodes. The second-order subgraph of a functional node includes the node itself (i.e., vertex), the adjacent nodes of the vertex, the adjacent nodes of the adjacent nodes, and the edges between these nodes. After determining a functional node, the first-order subgraph and multi-order subgraph of the functional stage can be obtained from the adjacency matrix of the matrix information.

[0078] The matrix information is input into the feature encoding network and the attention network respectively. The target network outputs a first prediction result after processing, and the attention network outputs a second prediction result. The first prediction result includes the learned adjacency matrix of the first-order subgraph and the second-order subgraph of each functional node, and the second prediction result includes the low-dimensional node representation of each functional node.

[0079] The setting of the optimization function of the multi-order subgraph information bottleneck can be flexibly selected. For example, it can be any loss function that can learn the difference between the real features and minimize the difference, such as KL divergence, cross entropy loss function, etc., and its implementation method is not restricted.

[0080] In order to make the feature encoding network learn the minimum sufficient features of each functional node by maximizing the mutual information between the node features and the target, while constraining the mutual information between the node features and the matrix information. and its subgraphs , define the multi-order subgraph information bottleneck optimization problem as: ,in and Respectively represent the learned adjacency matrix and subgraph learned by the feature encoding network The real adjacency matrix of By subgraph The predicted connection relationships between functional nodes in are directly described as learned self-supervisory labels. In addition, Elements Expressed as ,in , , is the low-dimensional node representation of the functional node, Represents the Sigmoid function.

[0081] In order to enable the feature encoding network to effectively learn node topology information at multiple scales, in the subgraph information bottleneck optimization problem, the embodiment of the present invention introduces a multi-order subgraph to effectively learn node topology information at multiple scales. Second-order subgraph as central node Defined as , where the node set of the second-order subgraph is defined as , and the edge set of the second-order subgraph is defined as: , The edge set that represents the topological graph, A collection of nodes that represents a topology graph.

[0082] Furthermore, assuming that a given function node The first-order subgraph of and the second-order subgraph , the multi-order subgraph information bottleneck optimization problem is reformulated as: , Characterization Second-order subgraph of functional nodes and learning adjacency matrix The mutual information of Characterization First-order subgraph of functional nodes and learning adjacency matrix The mutual information of Characterizing the hyperparameters of the feature encoding network, first-order subgraph has the same definition as a general subgraph, and They represent the feature encoding networks from and The learned adjacency matrix.

[0083] Since the mutual information between two variables in high-dimensional space cannot be directly calculated, variational inference is used to approximate the mutual information. , is the true posteriori The variational approximation of , so that: , Represents entropy.

[0084] From this, it can be deduced that The lower bound of is: , Characterize the topology diagram.

[0085] set up To represent the true prior distribution using variational approximation The variational approximation of , then the upper bound of the second term of the multi-node subgraph information bottleneck optimization problem expression can be calculated as: .

[0086] Combining the lower bound of the first term and the upper bound of the second term, we can get the overall lower bound of the multi-order topological information bottleneck optimization problem, which is expressed as: .

[0087] As mentioned above, the first term can be calculated by the variable cross entropy function, and the second term is quantified by the Kullback-Leibler divergence (i.e., KL divergence) between the true posterior distribution and the prior distribution. Therefore, the loss function of the multi-order topological information bottleneck optimization problem, that is, the optimization function of the multi-order subgraph information bottleneck, includes the variable cross entropy function and the KL divergence function. At this time, refer to Figure 6 The process of obtaining the information bottleneck optimization value in step 252 may include steps 252-1 to 252-5.

[0088] Step 252 - 1 , according to the learned adjacency matrix and the real adjacency matrix of the first-order subgraph of each functional node, the KL divergence function is used to obtain the first-order distribution difference value.

[0089] Step 252 - 3 , according to the learned adjacency matrix and the real adjacency matrix of the second-order subgraph of each functional node, a variable cross entropy function is used to obtain the second-order distribution difference value.

[0090] Step 252 - 5 , combining the first-order distribution difference value and the second-order distribution difference value to obtain the information bottleneck optimization value.

[0091] At this time, the process of obtaining the information bottleneck optimization value from step 252-1 to step 252-5 can be expressed by a formula, that is, the optimization function includes: .

[0092] in, represents the total number of functional nodes of the sample service network, Characterization The learned adjacency matrix of the second-order subgraph of function nodes, Characterization The true adjacency matrix of the second-order subgraph of function nodes, represents the difference value of the second-order distribution, represents the hyperparameters of the feature encoding network, Characterization The learned adjacency matrix of the first-order subgraph of function nodes, Characterization The true adjacency matrix of the first-order subgraph of function nodes, Represents the first-order distribution difference value.

[0093] At the same time, the setting of the global topological constraint function can be flexibly selected. For example, it can be a mean square error loss function, a KL divergence, a cross entropy loss function, or any other function that can measure the difference between the second prediction result output by the attention network and the actual connection relationship. The implementation method is not restricted.

[0094] When the second prediction result includes a low-dimensional node representation of each functional node, that is, the attention network outputs a low-dimensional node representation of each functional node, a global topological constraint function is constructed by combining graph variational inference and a function that measures the difference between the learned connection relationship and the actual connection relationship.

[0095] Therefore, by minimizing the difference between the approximate distribution of the connection relationship and the true posterior distribution through graph variational inference, and measuring the difference between the learned connection relationship and the true connection relationship, the information capture of the global topology and the local topology in the subgraph information bottleneck learning process is balanced. On this basis, referring to Figure 7 , the process of obtaining the global topology loss value in step 253 includes steps 253-1 to 253-7.

[0096] Step 253 - 1 , obtaining the learning connection relationship between each functional node in the sample service network according to each low-dimensional node representation.

[0097] Step 253 - 3 , obtaining a connection difference value according to the learned connection relationship and the real connection relationship.

[0098] Step 253-5, the step obtains the connection distribution difference value based on the low-dimensional node representation and graph variational inference.

[0099] Step 253 - 7 , obtaining a global topology loss value according to the connection difference value and the connection distribution difference value.

[0100] By way of example, in step 253 - 1 , any two low-dimensional node representations are input into an interior point function or an MLP model to obtain a connection relationship between functional nodes corresponding to the two low-dimensional nodes output by the interior point function or the MLP model.

[0101] Among them, the learning connection relationship between two functional nodes can be expressed as: , Characterization The low-dimensional node representation of the functional nodes, It can be an interior point function or an MLP model. Characterizing low-dimensional node representations Corresponding function node and low-dimensional node representation Corresponding function node The learning connection relationship between them.

[0102] Furthermore, in step 253-3, any difference measurement function may be used to calculate the connection difference value between the learned connection relationship and the real connection relationship, for example, the L1 norm, the L2 norm, or any other loss function.

[0103] When using the L2 norm, the learned connection relationship and the true connection relationship between any two functional nodes can be expressed as: ,in, Characterizing functional nodes and function nodes The real connection between Characterizing functional nodes and function nodes The learning connection relationship between them.

[0104] At the same time, refer to Figure 8 In step 253-5, the connection distribution difference value is obtained by following the steps 25-51 to 25-55, that is, the difference between the approximate distribution of the connection relationship and the true posterior distribution is minimized through graph variational inference.

[0105] Step 25-51, calculate the mean and standard deviation of each low-dimensional node representation, and obtain the variational distribution of each low-dimensional node representation based on the mean and standard deviation.

[0106] Step 25-53, obtain the forward propagation result based on the low-dimensional node representation, feature matrix and multi-layer perception.

[0107] Step 25-55, obtain the connection distribution difference value according to the variational distribution, the forward propagation result and the KL divergence function.

[0108] The process from step 25-51 to step 25-55 can be expressed by the formula: ,in, , Indicates A low-dimensional node representation, Characterize the feature matrix, Characterize the forward propagation model, and Represent the mean and standard deviation of the low-dimensional node representation respectively.

[0109] Therefore, the calculation formula of the global topology loss value, that is, the global topology constraint function can be expressed as: .

[0110] Then, the joint function can be expressed as: .

[0111] Through the above steps 21 to 25, i.e., the corresponding sub-steps, a feature encoding model is trained that can not only effectively learn node topology information at multiple scales, but also balance the information capture of global topology and local topology in the subgraph information bottleneck learning process. This enables the feature encoding model to combine the subgraph multi-level node semantic information (i.e., feature matrix) and multi-scale topology information (i.e., adjacency matrix) from the perspective of information theory to perform feature mining and obtain network node features with multi-scale constraint information in latent space, so as to obtain more accurate node features of each functional node in the service network to be tested (node ​​features of a single functional node).

[0112] After the node features of each functional node in the service network to be tested are obtained by using the trained feature coding model, the method of obtaining the node identification result in step 15 can be flexibly selected. For example, classification can be performed according to preset rules, or classification can be performed using a pre-trained classification model to obtain the classification results of each functional node. The above methods are only examples, and their implementation methods are not limited.

[0113] For example, the node features are input into the support vector machine to obtain the classification results of each functional node in the service network to be tested. That is, the graph convolutional neural network is trained according to the gradient descent method, etc., and the end-to-end network training is completed to obtain the feature encoding model. The feature encoding model is used to obtain the final node features of the service network function node, and the support vector machine is used to classify and identify the network function node based on the final node features of the service network function node.

[0114] Based on the same concept as the above-mentioned service network function node identification method based on subgraph information bottleneck, refer to Fig. 9 The embodiment of the present invention further provides a service network function node identification device 30, including a network modeling module 310, a feature extraction module 320 and a node identification module 330.

[0115] The network modeling module 310 is used to obtain a topology map of the service network to be tested, and obtain a feature matrix and an adjacency matrix of the topology map, wherein the topology map includes a plurality of functional nodes.

[0116] The feature extraction module 320 is used to input the feature matrix and the adjacency matrix into the feature coding model, and use the feature coding model to obtain the node features of the service network to be tested according to the feature matrix and the adjacency matrix. The feature coding model is a model obtained by training the optimization function of the joint multi-order subgraph information bottleneck and the global topology constraint function.

[0117] The node identification module 330 is used to obtain the node identification result of the service network to be tested according to the node characteristics.

[0118] The network modeling module 310, feature extraction module 320 and node identification module 330 can be deployed in Figure 1 The service network function node identification system 10 is shown in the identification system.

[0119] By way of example, a sample processing module and a model training module may also be included to obtain a feature encoding model.

[0120] The sample processing module is used to obtain the matrix information and real features of the sample service network, as well as the real connection relationship between the functional nodes in the sample service network; and use the real features and real connection relationship as labels for the matrix information. The matrix information includes the adjacency matrix and the feature matrix.

[0121] The model training module is used to use the matrix information as the input of the target network and perform iterative training using the joint function to obtain the feature encoding model.

[0122] The target network includes a feature encoding network for learning the features of each functional node and an attention network for learning the connection relationship between functional nodes. The joint function includes an optimization function of a multi-order subgraph information bottleneck and a global topology constraint function. The global topology constraint function is to calculate the loss value between the real connection relationship and the output result of the attention network, and the optimization function is to calculate the loss value between the real feature and the output result of the feature encoding network.

[0123] The above-mentioned service network functional node identification device 30, under the action of the network modeling module 310, the feature extraction module 320 and the node identification module 330, in a data-driven manner, combines the multi-level node semantic information and multi-scale topological information of the subgraph from the perspective of information theory to obtain network node features with multi-scale constraint information in the latent space, and can capture more accurate functional node features in the service network, thereby greatly improving the accuracy of service network node identification.

[0124] For the specific implementation and effect of the service network function node identification device 30, please refer to the description of the implementation of the service network function node identification method above, such as, for the specific implementation and effect of the network modeling module 310, please refer to the description of the relevant content of step 11 above, for the specific implementation and effect of the feature extraction module 320, please refer to the description of the relevant content of step 13 above, for the specific implementation and effect of the node identification module 330, please refer to the description of the relevant content of step 15 above, and no further details will be given here.

[0125] In addition, each module of the above-mentioned service network function node identification device 30 can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor 220 in the electronic device 20 in the form of hardware, or can be stored in the memory 210 of the electronic device 20 in the form of software, so that the processor 220 calls and executes the operations corresponding to the above-mentioned modules to implement the service network function node identification method provided above.

[0126] An embodiment of the present invention also provides an electronic device 20, including a processor 220 and a memory 210, wherein the memory 210 stores a computer program executable by the processor 220, and the processor 220 can execute the computer program to implement the service network function node identification method provided above.

[0127] The embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by the processor 220, the service network function node identification method proposed in the embodiment of the present invention is implemented.

[0128] In several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of the code, and a module, a program segment or a part of the code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0129] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0130] If the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory 210 (ROM, Read-Only Memory), random access memory 210 (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0131] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying service network function nodes based on subgraph information bottleneck, characterized in that: The method comprises: Obtaining a topological map of the service network to be tested, and obtaining a characteristic matrix and an adjacency matrix of the topological map; wherein the topological map includes a plurality of functional nodes; Inputting the feature matrix and the adjacency matrix into a feature coding model, and using the feature coding model to obtain node features of the service network to be tested according to the feature matrix and the adjacency matrix; wherein the feature coding model is a model obtained by training the optimization function of the joint multi-order subgraph information bottleneck and the global topology constraint function; A node identification result of the service network to be tested is obtained according to the node characteristics.

2. The method for identifying service network function nodes based on subgraph information bottleneck according to claim 1, characterized in that: The step of obtaining the feature encoding model comprises: Obtaining matrix information and real features of the sample service network, as well as real connection relationships between functional nodes in the sample service network; wherein the matrix information includes an adjacency matrix and a feature matrix; Using the real features and the real connection relationship as labels of the matrix information; The matrix information is used as the input of the target network, and iterative training is performed using a joint function to obtain a feature encoding model; wherein the target network includes a feature encoding network for learning the features of each functional node and an attention network for learning the connection relationship between functional nodes; the joint function includes an optimization function of a multi-order subgraph information bottleneck and a global topology constraint function, the global topology constraint function is used to calculate the loss value between the real connection relationship and the output result of the attention network, and the optimization function is used to calculate the loss value between the real feature and the output result of the feature encoding network.

3. The method for identifying service network function nodes based on subgraph information bottleneck according to claim 2 is characterized in that: The step of using the matrix information as the input of the target network and performing iterative training using a joint function to obtain a feature encoding model includes: Inputting the matrix information into the feature encoding network and the attention network respectively, obtaining a first prediction result output by the target network and a second prediction result output by the attention network; Based on the real connection relationship and the second prediction result, the global topology constraint function is used to obtain a global topology loss value; Based on the real feature and the first prediction result, using the optimization function to obtain an information bottleneck optimization value; The global topology loss value and the information bottleneck optimization value are combined to obtain a joint loss value; According to the joint loss value and back propagation, updating the parameters of the feature encoding network and the attention network; When the joint loss value converges, using the current feature encoding network as a feature encoding model; When the joint loss value does not converge, return to the step of inputting the matrix information into the feature encoding network and the attention network respectively to obtain a first prediction result output by the target network and a second prediction result output by the attention network.

4. The method for identifying service network function nodes based on subgraph information bottleneck according to claim 3 is characterized in that: The optimization function includes a variable cross entropy function and a KL divergence function, the real feature includes a real adjacency matrix of a first-order subgraph and a second-order subgraph of each functional node, and the first prediction result includes a learned adjacency matrix of a first-order subgraph and a second-order subgraph of each functional node; The step of obtaining the information bottleneck optimization value by using the optimization function based on the real feature and the first prediction result comprises: According to the learned adjacency matrix and the real adjacency matrix of the first-order subgraph of each functional node, the KL divergence function is used to obtain a first-order distribution difference value; According to the learned adjacency matrix and the real adjacency matrix of the second-order subgraph of each functional node, the variable cross entropy function is used to obtain a second-order distribution difference value; The information bottleneck optimization value is obtained by combining the first-order distribution difference value and the second-order distribution difference value.

5. The method for identifying service network function nodes based on subgraph information bottleneck according to claim 4 is characterized in that: The optimization function includes: ; in, represents the total number of functional nodes of the sample service network, Characterization The learned adjacency matrix of the second-order subgraph of function nodes, Characterization The true adjacency matrix of the second-order subgraph of function nodes, represents the difference value of the second-order distribution, represents the hyperparameters of the feature encoding network, Characterization The learned adjacency matrix of the first-order subgraph of function nodes, Characterization The true adjacency matrix of the first-order subgraph of function nodes, Represents the first-order distribution difference value.

6. The method for identifying service network function nodes based on subgraph information bottleneck according to claim 3 is characterized in that: The second prediction result includes a low-dimensional node representation of each functional node; The step of obtaining a global topology loss value based on the real connection relationship and the second prediction result by using the global topology constraint function comprises: Obtaining a learning connection relationship between each functional node in the sample service network according to each low-dimensional node representation; Obtaining a connection difference value according to the learned connection relationship and the real connection relationship; Based on the low-dimensional node representations and graph variational inference, obtaining a connection distribution difference value; A global topology loss value is obtained according to the connection difference value and the connection distribution difference value.

7. The method for identifying service network function nodes based on subgraph information bottleneck according to claim 6, characterized in that: The step of obtaining the connection distribution difference value based on each of the low-dimensional node representations and graph variational inference includes: Calculating the mean and standard deviation of each of the low-dimensional node representations, and obtaining the variational distribution of each of the low-dimensional node representations according to the mean and standard deviation; Obtaining forward propagation results according to the low-dimensional node representations, the feature matrix and the multi-layer perception; A connection distribution difference value is obtained according to the variational distribution, the forward propagation result and the KL divergence function.

8. The method for identifying service network function nodes based on subgraph information bottleneck according to claim 6, characterized in that: The step of obtaining the learning connection relationship between each functional node in the sample service network according to each low-dimensional node representation includes: Any two of the low-dimensional node representations are input into the interior point function to obtain the learning connection relationship between the functional nodes corresponding to the two low-dimensional nodes.

9. The method for identifying service network function nodes based on subgraph information bottleneck according to claim 7 or 8, characterized in that: The global topology constraint function includes: ; in, Represents a function node With function node The real connection between , Indicates A low-dimensional node representation, Characterize the interior point function, Represents a function node With function node The learning connection between , represents the total number of low-dimensional node representations, Characterize the feature matrix, Characterize the forward propagation model, and Represent the mean and standard deviation of the low-dimensional node representation respectively.

10. The method for identifying service network function nodes based on subgraph information bottleneck according to claim 1 or 2, characterized in that: The step of obtaining the node identification result of the service network to be tested according to the node characteristics includes: The node features are input into a support vector machine to obtain classification results of each functional node in the service network to be tested.

Images