Deep Clustering with Spherical Distance in Latent Space

This paper studies the problem of deep joint-clustering using auto-encoder. For this task, most algorithms solve a multi-objective optimization problem, where it is then transformed into a sing-objective problem by linear scalarization techniques. However, it introduces the scaling problem in latent space in a class of algorithms. We propose an extension to solve this problem by using scale invariance distance functions. The advantage of this extension is demonstrated for a particular case of joint-clustering with MSSC (minimizing sum-of-squares clustering). Numerical experiments on several benchmark datasets illustrate the superiority of our extension over state-of-the-art algorithms with respect to clustering accuracy.