CI-GAN : Co-Clustering By Information Maximizing Generative Adversarial Networks
Jaejun Lee, Hyun‐Chul Lee, Tomasz Palczewski · 2021
Simultaneously clustering rows and columns of a matrix, co-clustering can exploit the complex relationships between two different domains and identify groups of distinct nature. In this work, we introduce CI-GAN, a novel GAN-based approach for co-clustering. The model exploits two distinct GAN that cluster each domain independently and combines them intelligently by maximizing the mutual information between the input data and the generated co-clusters. From the experiments constructed with image, audio, and text datasets, it is found that such a systematic way of sharing information between the networks can improve co-clustering performance substantially; when CI-GAN is compared against five standard algorithms, it consistently reports the highest accuracy on both synthetic and real datasets.