Self-attention Adversarial Based Deep Subspace Clustering

Subspace clustering is a popular spectral clustering framework for high-dimensional data. In recent years, deep neural networks are witnessed to effectively mine the features of data, which has attracted the attention from various fields. Deep subspace clustering aims to learn the low-dimensional feature representation of data via the deep network, and subsequently calculate the similarity matrix fed into spectral clustering framework for final clustering result. However, real data are often with too high dimensions and complex data structure. How to obtain a robust data representation and improve clustering performance remains a challenge. Therefore, this paper proposes a deep subspace clustering algorithm (SAADSC) based on the self-attention adversarial mechanism. Specifically, the self-attention adversarial network is developed to impose a priori distribution constraint in the feature learning of the autoencoder to guide the learned feature representation to be more robust, thereby improving clustering accuracy. Through the experiments on multiple datasets, the results show that the proposed algorithm in this paper is superior to the state-of-the-art methods in terms of accuracy rate (ACC) and standard mutual information (NMI). Copyright ©2022 Acta Automatica Sinica. All rights reserved.