A Deep Semi-Supervised Community Detection Based on Point-Wise Mutual Information
Kamal Berahmand, Yuefeng Li, Yue Xu
Abstract
Network clustering is one of the fundamental unsupervised methods of knowledge discovery. Its goal is to group similar nodes together without supervision or prior knowledge of the nature of the clusters. Among various clustering methods, semi-supervised clustering detection is one of the most promising approaches for community detection because of its ability to employ side information to better understand network topology. However, most of the previous work faces two problems: the use of linear methods to reduce dimensionality and the random selection of side information, and as a result of these two drawbacks, semi-supervised community detection methods are less efficient. To fill these gaps, we developed an end-to-end deep semi-supervisor community detection (DSSC) for complex networks. A new learning objective is designed that uses a semi-autoencoder (SeAE) with a defined pair-wise constraint matrix based on point-wise mutual information (PMI) in the representation layer to accurately learn distinctive features and, in the clustering layer, adds a pair-wise constraint as a term to minimize distance within the cluster while the distance between clusters increases. The results show that our method performs unexpectedly well in comparison to the existing state-of-the-art community detection methods in complex networks.