Semi-supervised classification based on subspace sparse representation

• Published in: Knowledge and information systems Vol. 43; no. 1; pp. 81 - 101
• Main Authors: Yu, Guoxian, Zhang, Guoji, Zhang, Zili, Yu, Zhiwen, Deng, Lin
• Format: Journal Article
• Language: English
• Published: London Springer London 01.04.2015; Springer Nature B.V
• Subjects: Classification; Classifiers; Computer Science; Construction; Data Mining and Knowledge Discovery; Database Management; Discriminant analysis; Face recognition; Graph theory; Graphs; Information Storage and Retrieval; Information systems; Information Systems and Communication Service; Information Systems Applications (incl.Internet); Information technology; IT in Business; Methods; Neighborhoods; Regression; Regression analysis; Regular Paper; Representations; Sparsity; Studies; Subspaces; High-dimensional data; Subspaces sparse representation; Semi-supervised classification; Graph construction
• ISSN: 0219-1377; 0219-3116
• DOI: 10.1007/s10115-013-0702-2

Graph plays an important role in graph-based semi-supervised classification. However, due to noisy and redundant features in high-dimensional data, it is not a trivial job to construct a well-structured graph on high-dimensional samples. In this paper, we take advantage of sparse representation in random subspaces for graph construction and propose a method called Semi-Supervised Classification based on Subspace Sparse Representation , SSC-SSR in short. SSC-SSR first generates several random subspaces from the original space and then seeks sparse representation coefficients in these subspaces. Next, it trains semi-supervised linear classifiers on graphs that are constructed by these coefficients. Finally, it combines these classifiers into an ensemble classifier by minimizing a linear regression problem. Unlike traditional graph-based semi-supervised classification methods, the graphs of SSC-SSR are data-driven instead of man-made in advance. Empirical study on face images classification tasks demonstrates that SSC-SSR not only has superior recognition performance with respect to competitive methods, but also has wide ranges of effective input parameters.