Information Theoretic Learning-based Deep Embedded Clustering (ITL-DEC)

Clustering as the best grouping algorithm for the datasets is a main problem in various data-driven scientific and real-world applications. There are several methods based on some similarity measures for clustering, mostly suffering from high computational complexity on large-scale datasets. The performance of clustering methods highly depends on the quality of data representation; thus, in the literature, various linear and nonlinear representation methods and deep learning-based clustering algorithms have been exploited. This paper presents a novel fully unsupervised deep clustering method, namely information theoretic learning-based deep embedded clustering (ITL-DEC), with end-to-end training capable of simultaneously learning both feature representations and cluster assignments using deep neural networks (DNN). We use autoencoder as our powerful feature extraction deep neural network and two information-theoretic divergence measures, Cauchy-Schwarz divergence and Jensen-Shannon divergence, as cost functions to train the network parameter and appropriate clustering feature space. Experiments performed on three different benchmark datasets validate the effectiveness of the proposed method.