Improvement the Community Detection with Graph Autoencoder in Social Network Using Correlation-Based Feature Selection Method

Abstract

Background:
 In this paper, we aim to improve community detection methods using Graph Autoencoder. Community detection is a crucial stage in comprehend the purpose and composition of social networks.
 Materials and Methods:
 We propose a Community Detection framework using the Graph Autoencoder (CDGAE) model, we combined the nodes feature with the network topology as input to our method. A centrality measurement-based strategy is used by CDGAE to deal with the featureless dataset by providing artificial attributes to its nodes. The performance of the model was improved by applying feature selection to node features
 The basic innovation of CDGAE is that added the number of communities counted using the Bethe Hessian Matrix in the bottleneck layer of the graph autoencoder (GAE) structure, to directly extract communities without using any clustering algorithms.
 Results:
 According to experimental findings, adding artificial features to the dataset's nodes improves performance. Additionally, the outcomes in community detection were much better with the feature selection method and a deeper model. Experimental evidence has shown that our approach outperforms existing algorithms.
 Conclusion:
 In this study, we suggest a community detection framework using graph autoencoder (CDMEC). In order to take advantage of GAE's ability to combine node features with the network topology, we add node features to the featureless graph nodes using centrality measurement. By applying the feature selection to the features of the nodes, the performance of the model has improved significantly, due to the elimination of data noise. Additionally, the inclusion of the number of communities in the bottleneck layer of the GAE structure allowed us to do away with clustering algorithms, which helped decrease the complexity time. deepening the model also improved the community detection. Because social media platforms are dynamic.