Complex network node influence sorting method combining graph embedding and graph sampling

Abstract

The invention discloses a complex network node influence sorting method combining graph embedding and graph sampling in the field of complex networks. The complex network node influence sorting method comprises the following steps: S1, synthesizing a training data set based on different network topology models; s2, initializing the model; s3, pre-training and updating model embedding layer parameters by utilizing sorting learning; s4, model evaluation layer parameters are updated through reinforcement learning training; s5, sampling a real network by adopting random walk; s6, performing node influence sorting on each sub-network by using the trained model; and S7, calculating the node influence sequence of the original network through multi-source ranking fusion. According to the method, the model embedding layer is pre-trained through sorting learning, so that inductive graph embedding is effectively realized; reinforcement learning is used as node disconnection value learning, so that the algorithm has the capability of searching for disconnection nodes according to state iteration; the large-scale network is sampled through random walk, and the prediction efficiency and accuracy of the large-scale network are optimized.

Description

Technical Field

[0001] The present invention relates to a complex network node influence ranking method combining graph embedding and graph sampling, and belongs to the fields of network science and artificial intelligence. Background Art

[0002] In the real world, there are many interconnected and complex systems, such as the Internet, transportation systems, financial systems, social relationships, ecosystems, etc., which can all be abstracted as complex networks. Research in network science shows that networks in the real world are not all random networks. On the contrary, they more often show the characteristics of power-law distribution. This characteristic makes the network deeply affected by a small number of influential nodes, and the activation / disconnection of these nodes will significantly improve / degrade certain network functions. Therefore, the study of node influence has great theoretical significance and application value in a variety of network applications.

[0003] The existing complex network node influence evaluation methods can be divided into algorithms based on neighborhood information, algorithms based on information propagation, algorithms based on topology, algorithms based on eigenvectors, and multi-feature fusion algorithms according to different technical means. Algorithms based on neighborhood information determine the importance of nodes by analyzing the neighboring nodes of nodes. Related methods include HD, CI, LC, etc. This type of algorithm does not utilize the global information of the network, so it may not be able to accurately evaluate some nodes with large global influence; algorithms based on information propagation simulate the information propagation process and evaluate the influence of nodes from the perspective of information propagation. Related methods include CC, BC, SC, etc. This type of algorithm has the problem of high computational complexity and is difficult to handle tasks efficiently; algorithms based on topology determine the influence of nodes by evaluating the position of nodes in the network topology. Related methods include K-Core, HC, HKC, etc. This type of algorithm is very susceptible to misjudgment due to the influence of network structure, etc., and is difficult to adapt to complex and diverse network structures; algorithms based on eigenvectors consider the different contributions of neighboring nodes of different qualities to the influence of nodes, and evaluate the influence of nodes by calculating eigenvectors. Related methods include EC, PageRank, LeaderRank, etc. This type of algorithm cannot achieve ideal results on networks with low connectivity; multi-feature fusion algorithms often combine structural information with node features in one or more professional fields to make multi-attribute decisions, thereby obtaining more comprehensive results. With the development of artificial intelligence, academia has begun to introduce neural networks into multi-feature fusion methods to guide models to better adapt to actual application scenarios. Initially derived from simple unsupervised feature fusion, it gradually evolved to the introduction of supervised learning with objective evaluation indicators. However, although these methods have made some progress, they still have obvious shortcomings in terms of model generalization ability and practicality, and need further improvement and perfection. Summary of the invention

[0004] In view of the defects existing in the prior art, the purpose of the present invention is to achieve the above purpose, and the technical solution adopted by the present invention is:

[0005] A complex network node influence ranking method combining graph embedding and graph sampling, characterized by comprising the following steps:

[0006] (1) Synthesize training datasets based on different network topology models;

[0007] (2) Initialize the model, including the embedding layer, sorting layer, and evaluation layer;

[0008] (3) Use ranking learning pre-training to update the embedding layer parameters of the model;

[0009] (4) Update the evaluation layer parameters of the model through reinforcement learning training;

[0010] (5) Using the random walk algorithm to sample the original large real network, several sub-networks are obtained;

[0011] (6) Use the trained model to predict the influence ranking of nodes in each sub-network;

[0012] (7) The influence ranking results of each node in the original large network are obtained through multi-source ranking fusion.

[0013] In the method, step (1) further comprises:

[0014] (1.1) Comprehensively consider the network scale and network topology, and generate network instances by adjusting the three parameters of network model, number of network nodes and average degree;

[0015] (1.2) Select several centrality indicators to construct the node attributes of the network;

[0016] (1.3) The PageRank algorithm, which has a balanced performance in terms of complexity and accuracy, is used to generate node ranking labels.

[0017] In this method, the step (2) further comprises:

[0018] (2.1) Set the model’s hyperparameters;

[0019] (2.2) Randomly set the model’s embedding layer, sorting layer, and evaluation layer parameters.

[0020] In this method, the step (3) further comprises:

[0021] (3.1) Construct training batches from the generated graph dataset through dynamic batching;

[0022] (3.2) Compare the true rankings pairwise to obtain the true preference between nodes;

[0023] (3.3) Pass the training batch into the model embedding layer to obtain the feature matrix of the training batch;

[0024] (3.4) Concatenate the feature vectors in the feature matrix in pairs, pass them into the sorting layer, and output the predicted preference.

[0025] In this method, step (4) further comprises:

[0026] (4.1) Obtain training batches from the synthetic dataset by fixed batching;

[0027] (4.2) The current state of the network and the node attribute matrix are passed into the model embedding layer with fixed parameters to obtain the corresponding feature matrix;

[0028] (4.3) The feature matrix is ​​passed into the evaluation layer to obtain the disconnection value of each node in the current state;

[0029] (4.4) Determine the node selected for the disconnection action based on the current state and the disconnection value under the current state;

[0030] (4.5) Store the old state, new state, disconnection value, and timely reward into short-term memory;

[0031] (4.6) Repeat steps 2 to 5 until the network is completely disconnected;

[0032] (4.7) Use short-term memory to implement backpropagation according to the Bellman formula and update the model evaluation layer.

[0033] In this method, the step (5) further comprises:

[0034] (5.1) Determine the sampling parameters according to the global properties of the network, use the DFS-BFS hybrid sampling method to sample the original network, and obtain several sub-networks.

[0035] In this method, step (6) further comprises:

[0036] (6.1) Complete a complete disassembly round for each sub-network and record the order in which the nodes are disconnected.

[0037] In this method, step (7) further comprises:

[0038] (7.1) Use the weighted Borda counting method to calculate the predicted ranking of the original network from the predicted rankings of all sub-networks obtained from step (6). BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1is a schematic diagram of a flow chart of an embodiment of the present invention; DETAILED DESCRIPTION

[0040] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar concepts, objects, elements, etc. or concepts, objects, elements, etc. with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0041] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the field to which the present invention belongs and related fields. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0042] In order to solve the above technical problems, the present invention is achieved through the following technical solutions:

[0043] The present invention provides a complex network node influence ranking method combining graph embedding and graph sampling, such as Figure 1 FIG. 1 is a flow chart of an embodiment of the method of the present invention, and the specific operation steps are as follows:

[0044] 1. Synthesize training data sets based on different network topology models;

[0045] 1) Generate network instances by adjusting the three parameters of network model, number of network nodes and average degree: use ErdosRenyi (ER), Small World (SW) and Barabasi Albert (BA) network models to generate random network, small world network and scale-free network respectively; set the average degree of the instance in a way that the average degree k satisfies the power law distribution; divide the number of nodes into six intervals, namely 30-50, 50-100, 100-200, 200-300, 300-400 and 400-500, and randomly generate 100 network instances in each degree interval;

[0046]

[0047] 2) DC, K-Core and HITS are selected to represent the neighborhood information, topological structure and feature vector of nodes in the network respectively, and a feature matrix is ​​constructed for each network;

[0048] 3) The PageRank algorithm, which has a balanced performance in complexity and accuracy, is selected as the original node influence ranking algorithm to obtain the node ranking label.

[0049] 2. Initialize the model, including the embedding layer, sorting layer, and evaluation layer;

[0050] 1) Set the hyperparameters of the model;

[0051] 2) Use Kaiming normal distribution to initialize the weights of the convolutional layer, and use standard normal distribution to initialize the weights of the fully connected layer. The bias term is usually initialized to 0;

[0052] 3. Use ranking learning pre-training to update the embedding layer parameters of the model.

[0053] 1) Randomly select several networks from the generated graph dataset until the sum of the number of nodes reaches a threshold, and use the adjacency matrices of these networks to form a new diagonal matrix as the adjacency matrix of the training batch. The feature vector matrices of these networks are directly concatenated with the true rankings to form the training batch data;

[0054] 2) Compare the true rankings of each node in each network in the training batch and derive the true preference;

[0055] 3) Pass the training batch into the model embedding layer to obtain the feature matrix of the training batch;

[0056] 4) According to the original network division in the training batch, the feature vectors in the feature matrix are concatenated two by two, passed into the sorting layer, and the predicted preference is output.

[0057] 4. Update the evaluation layer parameters of the model through reinforcement learning training;

[0058] 1) Construct training batches from the synthetic dataset by fixing the number of networks to 1;

[0059] 2) The current state of the network and the node attribute matrix are passed into the model embedding layer with fixed parameters to obtain the corresponding feature matrix;

[0060] 3) The feature matrix is ​​passed to the evaluation layer to obtain the disconnection value of each node in the current state, and the future cumulative reward for disconnecting the node in the current state;

[0061] 4) Determine the node selected for the disconnection action based on the current state and the disconnection value under the current state;

[0062] 5) Use the difference in ANC before and after disconnection as an immediate reward, and store the old state, new state, disconnection value, and immediate reward in short-term memory;

[0063]

[0064] 6) Repeat steps 2 to 5 until the network is completely disconnected;

[0065] 7) Use short-term memory to implement back propagation according to the Bellman formula and update the model evaluation layer.

[0066] 5. Determine the sampling parameters according to the global properties of the network, use the DFS-BFS hybrid sampling method to sample the original network, and obtain several sub-networks;

[0067] 6. Complete a complete disassembly round for each sub-network and record the order in which the nodes are disconnected;

[0068] 7. Use the weighted Borda counting method to calculate the predicted ranking of the original network from the predicted rankings of all sub-networks obtained from step (6).

[0069] It should be noted that, in this application, relational terms such as "first" and "second" are only used 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", "comprise" 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, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0070] The above description is only a specific implementation of the present application, so that those skilled in the art can understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest range consistent with the principles and novel features applied for herein.

Claims

1. A complex network node influence ranking method combining graph embedding and graph sampling, characterized in that: The following steps are involved: (1) Synthesize training datasets based on different network topology models; (2) Initialize the model; (3) Update the model embedding layer parameters using ranking learning pre-training; (4) Update the model evaluation layer parameters through reinforcement learning training; (5) Using random walks to sample real networks; (6) Use the trained model to rank the node influence of each sub-network; (7) Calculate the node influence ranking of the original network through multi-source ranking fusion.

2. The large network node influence ranking algorithm integrating deep learning and random walk as claimed in claim 1 is characterized in that: The training set synthesis described in step (1) comprises the following steps: (1.1) Comprehensively consider the network scale and network topology, and generate network instances by adjusting the three parameters of network model, number of network nodes and average degree; (1.2) Select several centrality indicators to construct the node attributes of the network; (1.3) The PageRank algorithm, which has a balanced performance in terms of complexity and accuracy, is used to generate node ranking labels.

3. The large network node influence ranking algorithm integrating deep learning and random walk as claimed in claim 1 is characterized in that: The initialization model described in step (2) comprises the following steps: (2.1) Set the model’s hyperparameters; (2.2) Randomly set the model’s embedding layer, sorting layer, and evaluation layer parameters.

4. The large-scale network node influence ranking algorithm integrating deep learning and random walk as claimed in claim 1 is characterized in that: The sorting learning pre-training described in step (3) includes the following steps: (3.1) construct training batches from the generated graph dataset through dynamic batching; (3.2) Compare the true rankings pairwise to obtain the true preference between nodes; (3.3) Pass the training batch into the model embedding layer to obtain the feature matrix of the training batch; (3.4) Concatenate the feature vectors in the feature matrix in pairs, pass them into the sorting layer, and output the predicted preference; (3.5) The true preference and predicted preference are passed into the binary cross entropy loss function and then back-propagated to update the embedding layer weight parameters.

5. The large network node influence ranking algorithm integrating deep learning and random walk as claimed in claim 1 is characterized in that: The reinforcement learning training described in step (4) includes the following steps: (4.1) construct training batches from synthetic datasets by fixed batching; (4.2) The current state of the network and the node attribute matrix are passed into the model embedding layer with fixed parameters to obtain the corresponding feature matrix; (4.3) The feature matrix is ​​passed into the evaluation layer to obtain the disconnection value of each node in the current state; (4.4) Determine the node selected for the disconnection action based on the current state and the disconnection value under the current state; (4.5) Store the old state, new state, disconnection value, and timely reward into short-term memory; (4.6) Repeat steps 2 to 5 until the network is completely disconnected; (4.7) Use short-term memory to implement backpropagation according to the Bellman formula and update the model evaluation layer.

6. The large network node influence ranking algorithm integrating deep learning and random walk as claimed in claim 1, characterized in that: The network sampling described in step (5) includes the following steps: (5.1) Determine the sampling parameters according to the global properties of the network, use the DFS-BFS hybrid sampling method to sample the original network, and obtain several sub-networks.

7. The large network node influence ranking algorithm integrating deep learning and random walk as claimed in claim 1 is characterized in that: The subnetwork prediction described in step (6) includes the following steps: (6.1) Complete a complete disassembly round for each sub-network and record the order in which the nodes are disconnected.

8. The large network node influence ranking algorithm integrating deep learning and random walk as claimed in claim 1 is characterized in that: The original network prediction described in step (7) includes the following steps: (7.1) Use the weighted Borda counting method to calculate the predicted ranking of the original network from the predicted rankings of all sub-networks obtained from step (6).

Images