Auto-weighted Multi-view Clustering with Unified Binary Representation and Deep Initialization
Résumé
Clustering, as an integral part of exploratory data analysis, has gained renewed attention due to the prevalence of multi-representational or multi-view real-world data. However, its application becomes increasingly challenging in the face of large and heterogeneous datasets. Notably, existing techniques aimed at enhancing computational efficiency often possess drawbacks, such as assigning equal or static weights to views and samples, limiting the utilization of common and complementary features. Additionally, many methods execute the clustering task with arbitrary initialization, neglecting the rich structure of the joint discrete representation. In response to these challenges, this chapter introduces a novel approach named “auto-weighted binary multi-view clustering via deep initialization” designed for large-scale multi-view clustering. Two primary scenarios guide our approach: first, the differentiation between views based on sample importance, utilizing a dynamic learning strategy for automatic weighting of views and samples. Second, in the context of initializing binary clustering, we leverage a new CNN feature and employ a low-dimensional binary embedding, capitalizing on the efficient capabilities of Fourier mapping. Our proposed approach simultaneously learns a joint discrete representation and conducts direct clustering through constrained binary matrix factorization, solving the optimization problem in a unified learning model. Experimental results on several challenging datasets showcase the effectiveness and superiority of our approach over state-of-the-art methods, measured in terms of accuracy, normalized mutual information, and purity.