Clustering Ensemble Meets Low-rank Tensor Approximation

This paper explores the problem of clustering ensemble, which aims to combine multiple base clusterings produce better performance than that individual one. The existing ensemble methods generally construct a co-association matrix, indicates pairwise similarity between samples, as weighted linear combination connective matrices from different clusterings, and resulting matrix is then adopted input an off-the-shelf algorithm, e.g., spectral clustering. However, may be dominated by poor in inferior performance. In this paper, we propose novel low-rank tensor approximation based method solve global perspective. Specifically, inspecting whether two samples are clustered identical cluster under derive coherent-link contains limited but highly reliable relationships samples. We stack form three-dimensional tensor, low-rankness property further explored propagate information producing refined matrix. formulate proposed convex constrained optimization it efficiently. Experimental results over 7 benchmark data sets show model achieves breakthrough performance, compared with 12 state-of-the-art methods. To best our knowledge, first work explore potential on fundamentally previous approaches. Last not least, only one parameter, can easily tuned.