Multi-Level Structural Contrastive Subspace Clustering Network

Abstract

Deep subspace clustering methods based on autoencoder (AE) have achieved impressive performance in various applications. However, these methods often place excessive reliance on the AE framework, which focuses primarily on pixel-level reconstruction while overlooking the structural information inherent in the data. To overcome this limitation, we propose a novel approach called the Multi-level Structural Contrastive Subspace Clustering Network (MSCSCN). Unlike traditional AE-based methods, MSCSCN departs from the AE paradigm and introduces multi-level contrastive prediction to improve feature learning. Specifically, MSCSCN integrates multi-level features from both original and augmented data within a self-expression learning process, enhancing the learned pairwise affinities. Additionally, we propose a structural contrastive loss, which strengthens cluster boundary discrimination by effectively utilizing pairwise affinities and structural information. Our experimental results on several benchmark datasets demonstrate that MSCSCN outperforms competitive deep subspace clustering methods, highlighting its superior capability in improving clustering performance and capturing the underlying structural information within the data.