Deep Clustering and Representation Learning with Geometric Structure Preservation

In this paper, we propose a novel Deep Clustering and Representation Learning(DCRL) framework for learning effective representation with local and global structure preservation and partitioning data into clusters where each cluster contains data points from a compact manifold. In our framework, the latent space is manipulated to separate data points from different manifolds with Clustering Loss as guidance. Motivated by the observation that the clustering-oriented loss may corrupt the geometric structure of the latent space, two structure-oriented losses Isometric Loss and Ranking Loss are proposed to preserve the intra-manifold local structure and inter-manifold global structure, respectively. Our experimental results on various datasets show that DCRL achieves performance comparable to current state-of-the-art deep clustering algorithms and exhibits far superior performance in downstream tasks, demonstrating the importance and effectiveness of preserving geometric structure both locally and globally.