Robust Convex Clustering with Spectral Analysis-based Feature Selection

Clustering is a fundamental problem in many scientific applications. Traditional methods, such as k-means, Gaussian mixture models, and hierarchical clustering, however, are beset by local minima, which are sometimes drastically suboptimal. Furthermore recently introduced convex relaxations of methods average the weight of cluster assignment, which may lead performance deterioration. To address these issues, this paper presents a novel approach for robust convex clustering. In contrast to previously considered algorithms, the formulation utilize spectral analysis-based feature selection for alternating between minimization algorithm and multipliers. Rather than focusing on local features and their consistencies, our method aims at extracting sufficient information about the structure of the target concept. The experimental results demonstrate the effectiveness of our method in a variety of contexts on a real-world dataset.