Multi-channel Social Circle Recognition Device and Method Based on Enhanced Network Contrast Constraint

Abstract

The present invention relates to the technical fields of artificial intelligence and social networks, and discloses a multi-channel social network circle recognition device and method based on enhanced network contrast constraints. The device includes an input module, a solution module, and an output module. The work of the input module is data preprocessing, which abstracts the multi-channel social network existing in the real world into a multi-layer network structure. The solution module includes a network generation model and a node representation model. Based on the node representation model, the low-dimensional representations of the learnable enhanced network and the original input network are calculated, and based on the two obtained low-dimensional representations, the loss function is solved, and the parameters of the network generation model and the node representation model are optimized in reverse, so that the device has higher accuracy and robustness. Finally, the obtained multi-layer network low-dimensional representation is imported into the output module, and the community division is obtained by using Kmeans. In essence, the obtained community division corresponds to a social circle formed by a group of network user accounts in the real world.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and social networks, and more particularly to a multi-channel social circle recognition device and method based on enhanced network contrast constraints. Background Art

[0002] A social circle is a set of users who have common interests and are closely connected. With the development of the informationization process and the popularization of the network, the application analysis of user social circles in social networks has become increasingly important in different fields. The recognition of social circles has become a research focus in various fields. According to the inherent attributes of social circles (users within the same social circle have relatively close connections due to common interests; while users in different social circles have sparse connections due to fewer common topics), due to the intricate social relationships, abstracting the social system into the form of a complex network can effectively realize the processing, mining, and modeling of the social system. Specifically, the objects representing user identities in the social system (such as user accounts on social platforms) are abstracted as nodes in the network, and the connections between users (chatting, following common topics, mutual commenting, etc.) are mapped to the edges between nodes. Since the social activities of users in the real world are diverse, modeling only the single social relationship between users is difficult to truly reflect the social behavior of users. In real life, users often interact based on different social platforms and have multiple virtual identities in the network space. Therefore, information can be transmitted through different channels in the network. For example, users usually use WeChat to communicate with family and friends, and use software such as Weibo to communicate with netizens who are strangers but have common topic interests. These social relationships have different importance and meanings. The diversification of user relationships in multi-channel social networks makes it difficult for traditional community mining methods based on single-layer networks to accurately and effectively identify circles in social networks.

[0003] Therefore, the prior art abstracts multi-channel social networks into the form of multi-layer networks. Currently, scholars have proposed a large number of multi-layer network community mining methods for the special structure of multi-layer networks. Among them, some existing studies believe that only considering the topological structure of multi-layer networks is difficult to ensure the accuracy of the method in different network environments, and have proposed methods based on semi-supervised learning. However, due to the characteristics of large user volume and complex connection relationships in social networks, manually annotating data will cause huge economic and time costs. In view of this, the existing semi-supervised learning community mining methods are difficult to effectively solve the problem of social circle recognition. Summary of the Invention

[0004] Aiming at the problem that the prior art is difficult to effectively solve the problem of social circle recognition, the purpose of the present invention is to provide a multi-channel social circle recognition device and method based on enhanced network contrast constraints for accurately mining social circles in multi-channel social networks.

[0005] To achieve the above object, on the one hand, the present invention relates to a multi-channel social circle recognition device based on enhanced network contrast constraints, and adopts the following technical solutions:

[0006] A multi-channel social circle recognition device based on enhanced network contrast constraints, comprising:

[0007] An input module, which is used to read the input multi-channel social network and convert it into the form of an adjacency matrix and a feature matrix according to the format required by the algorithm;

[0008] A solution module, which includes a network generation model and a node representation model. Among them, the network generation model is used to aggregate multi-layer network information according to the input network and generate a learnable enhanced network; the node representation model is used to solve the low-dimensional representation of the input network, calculate the contrast loss according to the obtained low-dimensional representation, and then reversely optimize all model parameters in the device;

[0009] An output module, which based on the clustering algorithm, converts the multi-layer network consensus low-dimensional representation into the consensus community division of the multi-channel social network, that is, the social circle.

[0010] Furthermore, the multi-channel social circle recognition device is applicable to feature networks and non-feature networks. In a feature network, the multi-channel social circle recognition device can use the existing features as the feature matrix for solution; while in a non-feature network, the adjacency matrix can be used as the feature matrix for solution.

[0011] On the other hand, the present invention relates to a multi-channel social circle recognition method based on enhanced network contrast constraints, including the following steps:

[0012] S1: Model the multi-channel social network as a multi-layer network, and input the adjacency matrix of the multi-layer network representing the multi-channel social network into the multi-channel social circle recognition device; where each layer of the network represents an information dissemination channel composed of a social network platform, including the same user set and different social relationships. Each node in the network is a mapping of a user account in the social network platform in the real world, and the connection between nodes represents the social behavior between user accounts, and the social behavior includes forwarding and commenting;

[0013] S2: Set the number of iterations Epoch, the learning rate lr, and the dropout probability dp;

[0014] S3: The multi-channel social circle recognition device first constructs a network generation model based on the graph convolutional network (GCN) and the multi-layer perceptron (MLP), and constructs a node representation model based on the GCN. Essentially, the network generation model is a process of aggregating multi-layer network information to generate enhanced information; while the node representation model is a process of aggregating the enhanced network and the original multi-layer network information, and solving the contrastive loss based on the contrastive learning idea for the above results.

[0015] S4: Generate a learnable enhanced network based on the network generation model: First, use the GCN to extract and aggregate the information of each layer of the network, and generate the weights between each node and the rest of the nodes based on the MLP to generate the adjacency matrix A of the enhanced network. ′ ; The enhanced network is essentially an N×N matrix, where each vector represents the feature of each node generated after aggregating the multi-layer network information.

[0016] S5: Based on the node representation model, first calculate the low-dimensional representation of the learnable enhanced network generated in step S4, and then calculate the low-dimensional representation of each layer of the input multi-layer network.

[0017] S6: Use the low-dimensional representation of the learnable enhanced network and the low-dimensional representation of the multi-layer network to construct a contrast, and calculate the contrastive loss.

[0018] S7: Based on the loss function obtained in S6, reverse-optimize the network generation model and the node representation model.

[0019] S8: Loop and execute the training process S4 - S7, iteratively optimize the network generation model and the node representation model until the number of iterations reaches the predetermined parameter Epoch.

[0020] S9: Based on the trained node representation model, solve the consensus low-dimensional representation of the multi-layer network, and use Kmeans to solve the consensus community division.

[0021] Further, in S1, the multi-layer network is represented as G = {G 1 , G 2 , …, G L}, and it is solved using the adjacency matrix A = {A 1 , A 2 , …, A L} of the multi-layer network. L is the number of layers of the network, the matrix size is N×N, and N represents the number of nodes, that is, the number of user accounts.

[0022] Further, in S2, set Epoch = 100, lr = 0.004, dp = 0.2.

[0023] Further, in S4, the calculation formula for the low-dimensional representation of the GCN is as follows:

[0024]

[0025] Among them, X represents the input feature, A is the adjacency matrix of the multi-layer network, I represents the identity matrix, is the degree matrix of W (0) and W (1) are the weight matrices of the first and second layers of GCN respectively, and σ is the activation function.

[0026] Furthermore, in S5, the low-dimensional representation H' of the learnable enhancement network is normalized and denoted as Z'; the low-dimensional representation H of each layer in the multi-layer network i is normalized and denoted as Z i where i = {1, 2, … L}, and L is the number of network layers.

[0027] Furthermore, in S6, the loss function is solved based on the following formula:

[0028]

[0029] where diag(*) represents the diagonal matrix of calculating *, s() is the softmax function, τ represents a constant, mean() represents calculating the average value, and Z i T represents the transpose of Z i .

[0030] In a preferred embodiment of the present invention, τ = 0.6 is set.

[0031] Furthermore, in S9, each community obtained by solving is composed of several nodes, and each node represents a user account in the social network. The mined community structure essentially represents a social circle composed of a group of user accounts with the same hobbies and topics in reality.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] (1) The present invention abstracts a multi-channel social network into a multi-layer network form, where each layer of the network represents an information communication channel (i.e., depicts a social relationship under a certain social platform), solving the problem that the diversification of user relationships in a multi-channel social network makes it difficult for traditional community mining methods based on a single-layer network to accurately and effectively identify circles in the social network; by mining and aggregating multi-layer network information to generate a learnable enhanced network, and in addition, the network generation model parameters are synchronously iteratively updated according to the loss function, ensuring the effectiveness of the learnable enhanced network, improving the learning ability of the model, and thus effectively improving the accuracy and robustness of the device, and being able to accurately mine the consensus community structure of the multi-layer network in different network environments without the guidance of artificial labels.

[0034] (2) The device and method of the present invention are based on a node representation model, simultaneously solving the low-dimensional representations of the learnable enhanced network and the original input multi-layer network, and calculating a contrast loss based on the idea of contrastive learning, which can give full play to the guiding role of the learnable enhanced network.

[0035] (3) Comparing the present invention with other methods through comparative experiments on artificial datasets and real datasets of different scales, the method of the present invention has high accuracy and robustness, proving that the social circles in the multi-channel social network can be effectively mined through the method of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Other features, objectives, and advantages of the present invention will become more apparent by reading the detailed description of the non-limiting embodiments with reference to the following drawings:

[0037] Figure 1 is a flowchart of the method of the present invention;

[0038] Figure 2 is a mapping schematic diagram of a multi-channel social system, a multi-layer network, and social circles in reality;

[0039] Figure 3 is a schematic diagram when the device of the present invention is in use;

[0040] Figure 4 is an experimental graph for robustness analysis;

[0041] Figure 5 is an experimental graph for convergence analysis. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that those of ordinary skill in the art can make several deformations and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0043] A multi-channel social circle recognition device based on enhanced network contrast constraints, comprising:

[0044] An input module, which is used to read the input multi-channel social network and convert it into the form of an adjacency matrix and a feature matrix according to the format required by the algorithm;

[0045] A solution module, which includes a network generation model and a node representation model. Among them, the network generation model is used to aggregate multi-layer network information according to the input network and generate a learnable enhanced network; the node representation model is used to solve the low-dimensional representation of the input network, calculate the contrast loss according to the obtained low-dimensional representation, and then reversely optimize all model parameters in the device;

[0046] An output module, which is based on a clustering algorithm to convert the multi-layer network consensus low-dimensional representation into a consensus community division of the multi-channel social network, that is, a social circle.

[0047] A multi-channel social network circle recognition method based on enhanced network contrast constraints. First, a network generation model and a node representation model are constructed, and a learnable enhanced network of the multi-layer network is solved based on the network generation model. Then, a contrast is constructed based on the learnable enhanced network and the multi-layer network, and the parameters of the network generation model and the node representation model are updated based on the contrast loss.

[0048] As Figures 1 to 3 shown, it specifically includes the following steps:

[0049] S1: Model the multi-channel social network as a multi-layer network, where each layer of the network corresponds to a social channel (i.e., a social platform), the nodes represent the accounts of the corresponding users under different network platforms, and the edges between the nodes represent the social behaviors (such as mutual comments, chatting, etc.) between user accounts; the input represents the adjacency matrix A = {A 1 , A 2 , …, A L} of the multi-layer network G = {G 1 , A 2 , …, A L} representing the multi-channel social network, L is the number of layers of the network, the matrix size is N×N, N represents the number of nodes, that is, the number of user accounts, represents the weight of the social relationship between node i and node j in the real world (such as the number of mutual comments, the number of mutual chats, the number of forwarded messages, etc.). As Figure 2As shown, this example is a multi-channel social network composed of 6 social accounts. The two social platforms are QQ and WeChat respectively. In order to accurately depict the two social relationships, the device abstracts the social relationships between users based on the QQ and WeChat platforms as the 1st and 2nd layer networks respectively. The nodes in each layer network correspond one by one and all represent social users in the real world ( Figure 2 users V1-V6 in

[0050] S2: Set parameters such as the number of iterations Epoch, learning rate lr, and dropout probability dp. Among them, Epoch = 100, lr = 0.004, dp = 0.2;

[0051] S3: Build a network generation model based on Graph Convolutional Networks (GCN) and Multilayer Perceptron (MLP). Build a node representation model based on GCN. In order to distinguish the graph convolutional networks in different models, Figure 3 in it, GCN1 is used to represent the graph convolutional network module in the network generation model, and GCN2 is used to represent the graph convolutional network module in the node representation model;

[0052] S4: Based on the network generation model, calculate the low-dimensional representation of each layer network. The GCN calculation formula of this device is as follows:

[0053]

[0054] Among them, X represents the input feature, A is the adjacency matrix of the multi-layer network, I represents the identity matrix, is the degree matrix of W (0) and W (1) are the weight matrices of the first and second layer GCNs respectively, and σ is the activation function.

[0055] Due to the special hierarchical structure of the multi-layer network, in order to enable the enhanced network to have the information of each layer network, this device averages and aggregates the low-dimensional representations of each layer network to obtain the consensus low-dimensional representation of the multi-layer network as the input of the MLP. The weights between the nodes generated by the MLP are used as the learnable adjacency matrix A ′ ;

[0056] S5: Based on the node representation model, first calculate the low-dimensional representation H' of the learnable enhanced network, denoted as Z' after normalization, and then calculate the low-dimensional representation H i of each layer network in the input multi-layer network, denoted as Z i after normalization, where i = {1, 2,... L}, and L is the number of network layers;

[0057] S6: Based on the idea of contrastive learning, construct a contrast between the low-dimensional representation of the learnable enhancement network and the low-dimensional representation of the multi-layer network, and solve the loss function based on the following formula:

[0058]

[0059] where diag(*) represents calculating the diagonal matrix of *, s() is the softmax function, τ represents a constant, τ = 0.6 is set, mean() represents calculating the average value, and Z i T represents the transpose of Z i According to the formula described above, this formula calculates the normalized low-dimensional vectors Z' and Z to obtain the node similarity matrix between them, then solves the similarity between corresponding nodes through diag(), and obtains the average value of the similarity between nodes according to the mean() function, so as to obtain the overall similarity of the two low-dimensional representations; i

[0060] S7: Based on the loss function obtained in S6, reversely optimize the network generation model and the node representation model, so that the learnable enhancement network can better aggregate the information of the multi-layer network, while improving the robustness of the node representation model and ensuring its accuracy in different network environments;

[0061] S8: Loop and execute the training process S4 - S7 to iteratively optimize the network generation model and the node representation model until the number of iterations reaches the predetermined parameter Epoch;

[0062] S9: Based on the trained node representation model, solve the consensus low-dimensional representation of the multi-layer network, and use Kmeans to solve the consensus community division. Each community consists of several nodes, and each node represents a user account in the social network. Therefore, each community finally obtained by this device essentially represents a social circle composed of a group of social accounts in the input multi-channel social network. That is, in the example shown Figure 2 the proposed device divides 6 users into 2 social circles with different topics and hobbies.

[0063] To verify the device and method of the present invention, 7 different data sets are used, as shown in Table 1. Among them, the first five are real data sets, and Syn1 and Syn2 are artificial generated network groups for robustness testing. SND is a social network, and WTN and WBN are the World Trade Network and the Worm Brain Network respectively. Cora and Citeseer are both literature networks.

[0064] Table 1 Summary of Data Sets

[0065] Network Number of nodes Number of layers Number of communities SND 71 3 3 WTN 183 14 10 WBN 279 5 10 CoRA 1662 2 3 Citeseer 3312 2 3 Syn1 128 3 4-8 Syn2 128 9 4-8

[0066] ​The random network, real network, and learnable enhancement network were respectively used as the test objects, and Table 2 shows the results of the ablation experiment. The experimental results prove that in the real network CoRA and the artificial networks S1 and S2, the learnable enhancement network achieved the highest NMI. Especially in S2, compared with the random network, the improvement rate exceeded 100%. This experiment proves the effective improvement of the learnable enhancement network on the device performance.

[0067] Table 2 Results of the ablation experiment

[0068]

[0069] Figure 4 It is a graph for the robustness analysis experiment, including three comparison methods. Among them, NACC is the multi-channel social circle recognition device based on the enhanced network contrast constraint proposed by the present invention. Since the structure parameters of the artificially generated network are adjustable, in order to verify the robustness of the invented device, this experiment uses a large number of artificial networks instead of social networks for testing. The artificial networks control the network structure through μ, Dc, and Layer respectively. Among them, (a) and (b) belong to the first group of networks (Syn1), which includes 3-layer networks. The μ values of the two groups of networks are set to 0.5 and 0.6 respectively, and Dc = {0.2 - 0.8}. (c) and (d) belong to the network group Syn2, which includes 9-layer networks. The μ values are set to 0.5 and 0.6 respectively, and Dc = {0.2 - 0.8}. The comparison methods are SC-ML, COMCLUS, and MOEA-MultiNet respectively. It can be seen from the experimental results that in the dataset shown in Figure (a), the proposed NACC device has a large lead in accuracy, and under the networks with different Dc parameters, the fluctuation of NACC is significantly smaller than that of the comparison methods, indicating that under the guidance of the optimizable enhancement network, the device of the present invention can maintain a high accuracy when facing datasets with different topological structures.

[0070] To further prove the performance of the device of the present invention, μ is set to 0.6 (as shown in Figure (b)). At this time, the artificial network topology is complex, the node connections are relatively chaotic, and there is no obvious community relationship. In this type of dataset, NACC generally maintains a high accuracy. Especially when Dc = {0.3, 0.6}, the effect is significantly better than the comparative method. As shown in Figure (c), under the network parameters of μ = 0.5 and Layer = 9, NACC has a relatively obvious advantage. In the result shown in Figure (d), the NACC device has a poor effect in the Dc = 0.8 network. This is because Dc controls the degree difference of nodes between different layers of the artificial network, and the difference between different layers of this network is large, resulting in a poor effect when the invented device aggregates the low-dimensional representations of different network layers. In the network with Dc = {0.2 - 0.7}, the performance is significantly better than the comparative method, especially with a large lead when Dc = 0.3. Looking at the whole figure, the device proposed by the present invention has achieved the highest accuracy in nearly 89% of the datasets, and still has a relatively large advantage in the network with complex structure and fuzzy community (μ = 0.6). Experiments prove that this device can complete the multi-layer network community mining task, that is, the social circle recognition task, with high performance in networks of different structures.

[0071] The WBN real network and the 128-node artificial network are respectively used as test data to verify the changes in the accuracy (NMI) and loss function (Loss) of the device under the iteration times Epoch = [10, 30, 50, 70, 90, 110, 130, 150]. Figure 5 The results of the convergence analysis are shown. The left axis and the red hexagon broken line represent the change of NMI with the iteration times, and the right axis and the blue pentagram curve represent the change of the loss function value. In the WBN and artificial datasets (Synthetic), the loss function value (Loss) decreases regularly with the increase of the iteration times (Epoch), and the change rate is first fast and then slow, while the NMI value first increases and then decreases, reaching the peak at a certain specific Epoch value. This is because GCN, as the main component of the network generation model and node representation model in the NACC device, will cause the over-smooth problem during the process of increasing Epoch and network depth. In the WBN dataset, NMI reaches the highest at Epoch = 110, while in the artificial dataset, it reaches the optimal at Epoch = 50.

[0072] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various deformations or modifications within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A multi-channel social circle recognition method based on enhanced network contrast constraints, characterized in that, It includes the following steps: S1: Model the multi-channel social network as a multi-layer network and input the adjacency matrix representing the multi-layer network of the multi-channel social network into the multi-channel social circle recognition device; where each layer of the network represents an information dissemination channel composed of a social network platform, each node in the network is a mapping of user accounts in the social network platform in the real world, and the edges between nodes represent social behaviors between user accounts, and the social behaviors include forwarding and commenting; the multi-layer network is represented as G={G 1 , G 2 ,…,G L}, and use the adjacency matrix A={A 1 , A 2 ,…, A L} of the multi-layer network to solve, L is the number of layers of the network, the matrix size is , and N represents the number of nodes, that is, user accounts; S2: Set the number of iterations Epoch, learning rate lr, and dropout probability dp; S3: The multi-channel social circle recognition device first constructs a network generation model based on the graph convolutional network (GCN) and the multi-layer perceptron (MLP), and constructs a node representation model based on GCN; the network generation model is a process of aggregating multi-layer network information to generate enhanced information; while the node representation model is a process of aggregating the enhanced network and the original multi-layer network information, and solving the contrast loss based on the contrast learning idea for the result; S4: Generate a learnable enhanced network based on the network generation model: First, use GCN to extract and aggregate the information of each layer of the network, and generate the weights between each node and the remaining nodes based on MLP to generate the adjacency matrix of the enhanced network. ; The enhanced network is a matrix, where each vector represents the feature of each node generated after aggregating the information of multiple layers of the network. S5: Based on the node representation model, first calculate the low-dimensional representation of the learnable enhancement network generated in step S4, and then calculate the low-dimensional representation of each layer of the input multi-layer network; the low-dimensional representation of the learnable enhancement network , after normalization, is denoted as ; the low-dimensional representation H of each layer of the multi-layer network i , after normalization, is denoted as Z i , where i = {1, 2, …L}, and L is the number of network layers; S6: Use the low-dimensional representation of the learnable enhancement network to construct a contrast with the low-dimensional representation of the multi-layer network, calculate the contrast loss, and solve the loss function based on the following formula: ， where, diag(*) represents calculating the diagonal matrix of *, s() is the softmax function, τ represents a constant, mean() represents calculating the average value, and Z i T represents Z i is the transpose of; S7: Based on the loss function obtained in S6, reverse-optimize the network generation model and the node representation model; S8: Loop through the training process S4 - S7, iteratively optimize the network generation model and the node representation model until the number of iterations reaches the predetermined parameter Epoch; S9: Based on the trained node representation model, solve the consensus low-dimensional representation of the multi-layer network, and use Kmeans to solve the consensus community division; The multi-channel social circle recognition device includes: An input module, which is used to read the input multi-channel social network structure information and convert it into the form of an adjacency matrix and a feature matrix according to the format required by the algorithm; A solution module, which includes a network generation model and a node representation model. Among them, the network generation model is used to aggregate multi-layer network information according to the input network and generate a learnable enhancement network; the node representation model is used to solve the low-dimensional representation of the input network, calculate the contrast loss based on the obtained low-dimensional representation, and then reverse-optimize all model parameters in the device; An output module, which converts the consensus low-dimensional representation of the multi-layer network into the consensus community division of the multi-channel social network, that is, the social circle, based on the clustering algorithm.

2. The multi-channel social circle recognition method based on enhanced network contrast constraints according to claim 1, characterized in that, In S2, set Epoch = 100, lr = 0.004, dp = 0.

2.

3. The multi-channel social circle recognition method based on enhanced network contrast constraints according to claim 1, characterized in that, In S4, the calculation formula for the GCN low-dimensional representation is as follows: ， Among them, X represents the input feature, A is the adjacency matrix of the multi-layer network, , I represents the identity matrix, is the degree matrix of, , and are the weight matrices of the first and second layer GCNs respectively, and σ is the activation function.

4. The multi-channel social circle recognition method based on enhanced network contrast constraints according to claim 1, characterized in that, Set τ = 0.

6.

5. The multi-channel social circle recognition method based on enhanced network contrast constraints according to claim 1, characterized in that, In S9, each community obtained by solving is composed of several nodes, and each node represents a user account in the social network.

6. The multi-channel social circle recognition method based on enhanced network contrast constraints according to claim 1, characterized in that, The multi-channel social circle recognition device is applicable to both feature networks and non-feature networks. In a feature network, the multi-channel social circle recognition device uses the existing features as the feature matrix for solution; while in a non-feature network, the adjacency matrix is used as the feature matrix for solution.

Images