Multi-Head Self-Attention Network for Multi-View Deep Clustering

Multi-view clustering comprehensively processes data objects by studying different views or modalities, as well as the differences or complementary information between different views, then classifies the data objects into different groups by unsupervised learning. However, most of the work these days fuses the information of multiple views only after rough processing, which ignores the difference of information value among different data features in the process of deep fusion representation. In this paper, we propose a Multi-head Self-attention network for Deep Clustering (MSDC) to highlight valuable information features and improve the effect of multi-view clustering. Comprehensive experiments have been carried out on real world datasets and results show that MSDC outperforms other advanced baselines, demonstrating that MSDC provides an excellent approach for multi-view clustering.