Subspace Clustering Using Unsupervised Data Augmentation

Abstract

Subspace clustering is an unsupervised approach for determining the union of multiple subspaces that best fits a collection of high-dimensional samples. Self-expressive representation, that is, expressing samples as linear combination of other samples, is the core of most state-of-the-art subspace clustering approaches. However, existence of sufficiently well-spread samples within each subspace is crucial for precise representation, which might not always be available in real-world scenarios. Inspired by the remarkable influence of data augmentation on the performance of neural networks, we propose a scalable approach that employs data augmentation within subspace clustering. Benefiting from the increased diversity in data, we use augmented samples as an enlarged dictionary and combine the self-expressive representations based on the assumption that augmentation does not alter the labels of the samples. Significant improvement of the clustering performance on two real-world datasets demonstrates the effectiveness of the proposed approach.