Scalable multi-view discrete clustering with self-supervised constraints

Multi-view spectral clustering (MSC), which integrates information from diverse data perspectives to improve clustering accuracy, has been extensively studied and achieved remarkable performance. However, most existing MSC methods overlook the utilization of additional prior information, which limits the improvement of clustering performance. In addition, most MSC algorithms need to construct a full graph, which results in high computational cost and limits their application on large-scale datasets. Moreover, early methods adopt a two-step scheme involving relaxation followed by post-processing, which may lead to information loss and eventually produce sub-optimal results. To mitigate the problems outlined above, we propose a scalable multi-view discrete clustering with self-supervised constraints (SMDC-S²C) method, which is inspired by self-supervised clustering. SMDC-S²C has the following contributions: 1) an innovative yet straightforward approach is proposed to obtain high-confidence self-supervised information, which is defined as ensemble local cluster constraints and is applied to constrain the indicator matrix; 2) an enhanced coordinate descent (CD) method with simultaneous multi-row updates is developed to directly optimize the primal spectral problem, complemented by the introduction of the anchor graph technique; both make SMDC-S²C suitable for large-scale datasets; 3) the weight of each view is calculated automatically to measure their respective contributions. Extensive experiments across real-world datasets confirm the superiority of SMDC-S²C.