Triplets-based large-scale multi-view spectral clustering

Abstract

By integrating complementary information from multiple views to reach consensus, graph-based multi-view clustering describes data structure competently, thereby attracting considerable attention. It is time-bottlenecked by graph construction and eigen-decomposition in the big data era. Existing methods usually utilize anchor graph learning to address this issue. However, problems of unsupervised representative anchor selection, consensus or view-specific anchors, anchor alignment, etc., remain challenging. Moreover, excessive information is discarded for the sake of efficiency. Motivated by the essence of the anchor-based methods that utilizing representative point-to-point relations to reduce graph complexity, we generate triplets for each view based on neighborhood similarity to preserve point-to-point relations and local structure, and propose triplets-based large-scale multi-view spectral clustering (TLMSC). Subsequently, the triplet enhancement strategy is designed to select representative triplet relations to improve efficiency and clustering performance. Specifically, positive examples of triplets are filtered according to the view consensus to significantly increase the probability of positive examples belonging to the same cluster. The most indistinguishable hard negative examples are generated based on probabilities to improve discrimination performance. Guided by the enhanced triplets and its weights, an improved low-dimensional embedding is constructed through optimization, which further serves as an input to the proposed fast sparse spectral clustering (FSSC) to obtain clustering results. Numerous experiments validate the efficiency and superior performance of the proposed TLMSC. An average improvement of 12.25% at least in ACC compared to 10 state-of-the-art methods on 18 datasets. The code is available at github.com/ytccyw/TLMSC.