Adversarial domain adaptive subspace clustering
Mahdi Abavisani, Vishal M. Patel · 2018
We propose a novel method for clustering a collection of data that comes from several domains. Since members of the same class might look very different across different domains and because in a clustering problem we have no side information such as labels, the main challenge in a domain adaptive clustering problem is to group the data into different clusters regardless of their domain. We approach this problem by finding mappings that can transfer the data points between the domains. We use adversarial networks to approximate these mapping functions, and form a paired representation at each domain by mapping the data onto their counter domains. Finally, we employ a multimodal subspace clustering type algorithm to cluster the paired representations with respect to their subspaces. Various experiments on datasets with domain shifts show that our method performs significantly better than many competitive domain adaptive subspace clustering methods.