Spectral Ensemble Clustering with LDA-based Co-training for Multi-view Data Analysis

Abstract

Multi-view data has attracted more and more attention in recent years due to its practical significance. Generally, ensemble clustering and co-training are effective ways to analyze this kind of data. Among ensemble clustering algorithms, Spectral Ensemble Clustering (SEC) shows good performance by leveraging co-association matrix, which is constructed with base clustering result of single-view space. However, SEC is also limited by base clustering results. When the base clustering results of some views deviate greatly from the actual data distribution, it will affect the accuracy of co-association matrix, making it difficult for SEC to learn good cluster structure. Therefore, this paper proposes a novel algorithm for multi-view data clustering, namely Spectral Ensemble Clustering with LDA-based Co-training (LSEC). This method introduces a novel co-training strategy into ensemble clustering analysis. Specifically, based on the assumption that the clustering result of SEC has global cluster information, we use the class label obtained by SEC as reference label to perform Linear Discriminant Analysis (LDA) on original data to get refined data. Subsequently, K-means algorithm is performed on the refined data to get new base clustering results. The basic idea of this strategy is to use global cluster information to co-train local feature space. Experimental results on seven real world multi-view data sets show the effectiveness of proposed method.