Consecutive and Similarity Information Fused Graph Learning Using Semi-Nonnegative Matrix Factorization for Sequential Data Clustering

Abstract

Sequential data clustering plays an important role in many applications, such as motion recognition, video segmentation, gene sequence analysis, and so on. The key idea is to learn a locally continuous and connected representation by exploiting the similarity structure in the sequence to cluster them into a set of nonoverlap fragments. Most existing methods focus on exploring the local consecutive similarity structure but ignore other similarity structures, resulting in a barely satisfactory clustering performance. Given this gap in the research, this article focuses on fusing both the consecutive and similarity information in the representation learning process by exploiting different similarity structures in sequence, and a novel regularized semi-nonnegative matrix factorization (Semi-NMF)-based method for sequential data clustering is proposed. It presents two critical differences from other classic sequential data clustering methods: first, it learns a graph representation by using the nearest neighbors with the consecutive neighbor information in a one-stage Semi-NMF model. Second, two novel graph regularization terms are specifically designed to preserve both the consecutive relationship and similarity information simultaneously in the learning process. An iterative updating optimization algorithm is then proposed to solve the corresponding nonconvex optimization problem. Extensive clustering experiments on the real ordered face image and the sequential motion datasets demonstrate the effectiveness and superiority of the proposed method compared to other state-of-the-art methods.