关键词: 稀疏表示, 亲近分类, 流形学习, 鲁棒性

Abstract: As a Generalized Eigen-value based Local Information Proximal SVM（LIPSVM）aims at assigning data points to the closer of two nonparallel planes which are generated by their corresponding generalized eigen-value problems in LIPSVM. LIPSVM owns superiorities in both computation time and test correctness. Existing researches show that the performance of LIPSVM is sensitive to the parameters of model. To address this issue in LIPSVM, following the geometric intuition of LIPSVM, a robust classification method called sparse representation Proximal Support Vector Machine（PSVM） based on sparse representation technology is proposed. By exploring the discriminative sparse representation information among the training points, SPSVM not only keeps aforementioned characteristics of LIPSVM, but also has its additional advantages, e.g., robustness to noise data or outliers and avoiding the model parameters selection problem in LIPSVM. Experimental results on the artificial and benchmark datasets demonstrate the comparable learning performance of SPSVM with respect to several exiting methods.

Key words: sparse representation, proximal classification, manifold learning, robustness