Constrained Hard-Soft Subspace Clustering

In high-dimensional data, clusters are in subspaces of the original space. Traditional hard subspace search clustering algorithms selected dimensions, while soft projected algorithms assign weights to each dimension. In this paper, we propose a Constrained Hard-Soft Subspace Clustering algorithm which integrates hard subspace search and soft subspace weighting mechanism in one framework by introducing pairwise constraints. In the hard subspace search step, constraints are utilized to break the circular dependency by introducing must-link dimension relevance. In the soft subspace generation and data assignment step, we develop a novel optimization objective function which panelizes the constraint violation. Experimental results show that the proposed algorithm outperforms typical hard and soft subspace clustering algorithms in real world data sets, and its scalability is comparable to traditional subspace clustering algorithms.