Multi-view alternative hard c-means clustering

The ubiquity of multi-view data has attracted great attention for the development of multi-view clustering methods. However, the field still faces serious challenges. For example, noise and outliers are common in real-world multi-view data and tend to harm the accuracy of clustering methods. Additionally, many existing methods primarily focus on evaluating the whole contribution of each view, while neglecting the intra-view contributions. To address these issues, this paper introduces a robust multi-view alternative hard c-means clustering based on global weighting (MVAHCM-GW), which is an extension of the single-view alternative hard c-means. The unique aspect of MVAHCM-GW lies in its adoption of a non-Euclidean norm metric instead of the traditional Euclidean norm, thereby increasing its robustness against noise and outliers. Furthermore, we integrate view weights into the clustering process to better understand the role of each view. Expanding upon this, we also present a robust multi-view alternative hard c-means clustering based on local weighting (MVAHCM-LW). MVAHCM-LW allows for a more precise identification of the varying contributions of different clusters within each view, leading to improved clustering performance. We introduce a new product constraint on view weights for MVAHCM-GW and MVAHCM-LW, which has the advantage of not adding additional parameter. An optimization strategy has been developed for both MVAHCM-GW and MVAHCM-LW to achieve optimal clustering results. Experimental evaluations on benchmark datasets reveal that our proposed MVAHCM-GW and MVAHCM-LW methods surpass existing methods in performance, highlighting their effectiveness in contemporary multi-view data clustering scenarios.