Automatic Model Selection in Subspace Clustering via Triplet Relationships

• DOI: 10.1609/aaai.v32i1.11712
• Publication Date: 2022-06-24T21:08:34Z
• Authors: Jufeng Yang; Jie Liang; Kai Wang; Yong-Liang Yang; Ming-Ming Cheng
• Keywords: Robustness

This paper addresses both the model selection (i.e., estimating number of clusters K) and subspace clustering problems in a unified model. The real data always distribute on union low-dimensional sub-manifolds which are embedded high-dimensional ambient space. In this regard, state-of-the-art approaches firstly learn affinity among samples, followed by spectral to generate segmentation. However, arguably, intrinsic geometrical structures samples rarely considered optimization process. paper, we propose simultaneously estimate K segment according local similarity relationships derived from matrix. Given correlations define novel structure termed Triplet, each reflects high relevance locality three aimed be segmented into same subspace. While traditional pairwise distance can close between inter-cluster lying intersection two subspaces, wrong assignments avoided hyper-correlation proposed triplets due complementarity multiple constraints. Sequentially, greedily optimize new reward triplets. We fusion based similarities final Extensive experiments benchmark datasets demonstrate effectiveness robustness approach.