Fast Multi-view Clustering via Ensembles: Towards Scalability, Superiority, and Simplicity

Despite significant progress, there remain three limitations to the previous multi-view clustering algorithms. First, they often suffer from high computational complexity, restricting their feasibility for large-scale datasets. Second, typically fuse information via one-stage fusion, neglecting possibilities in multi-stage fusions. Third, dataset-specific hyperparameter-tuning is frequently required, further undermining practicability. In light of this, we propose a fast xmlns:xlink="http://www.w3.org/1999/xlink">m ulti-v xmlns:xlink="http://www.w3.org/1999/xlink">i ew xmlns:xlink="http://www.w3.org/1999/xlink">c lustering xmlns:xlink="http://www.w3.org/1999/xlink">e nsembles (FastMICE) approach. Particularly, concept random view groups presented capture versatile view-wise relationships, through which hybrid early-late fusion strategy designed enable efficient With multiple views extended xmlns:xlink="http://www.w3.org/1999/xlink">many groups, levels diversity (w.r.t. features, anchors, and neighbors, respectively) are jointly leveraged constructing view-sharing bipartite graphs early-stage fusion. Then, set diversified base clusterings different obtained fast graph partitioning, formulated into unified final late-stage Notably, FastMICE has almost linear time space free tuning. Experiments on 22 datasets demonstrate its advantages scalability (for extremely large datasets), superiority (in performance), simplicity (to be applied) over state-of-the-art. Code available: https://github.com/huangdonghere/FastMICE .