Subspace Clustering with A Hybrid Adaptive Graph Filter

Abstract

Subspace clustering is a powerful tool for grouping data samples into their underlying subspaces. In this paper, we propose an advanced subspace clustering algorithm called SCHAGF (Subspace Clustering with A Hybrid Adaptive Graph Filter). SCHAGF leverages the obtained reconstruction coefficient matrix to design a low-pass graph filter and a high-pass graph filter simultaneously. These graph filters are then integrated into a hybrid graph filter, which is used for designing a feature extraction function and a constraint for the reconstruction coefficient matrix. Then the hybrid graph filter and the coefficient matrix are iteratively updated to achieve optimal values. Our results demonstrate that the features extracted using the hybrid graph filter exhibit compactness within classes and discrimination between classes. Additionally, the new constraints significantly enhance the block-diagonal structure of the reconstruction coefficient matrix. Finally, plenty of subspace clustering experiments show that the SCHAGF outperforms the related algorithms. Moreover, by incorporating the thresholding technique, thresholding SCHAGF (TSCHAGF) is found to surpass some deep models.