Feature extraction using nonparametric margin discriminant analysis

High dimensional data often lies in a low dimension structure and is often recommended to perform dimensionality reduction before the tasks of classification or clustering. In this paper, we propose a new method called nonparametric margin discriminant analysis based on the trace ratio criterion (TRNMDA) to extract discriminant vectors from original high dimensional space. TRNMDA does not have the assumption on data distribution that overcomes the limitation of LDA, which can not perform well in non-gaussian data. Compared with other nonparametric methods like nonparametric maximum margin criterion (NMMC), TRNMDA takes full advantage of marginal information, which is essential for classification. Experiments conducted on three standard face data sets have validated that our proposed method is more effective than several state-of-the-art methods.