Experimental Study on Multiple LDA Classifier Combination for High Dimensional Data Classification

Multiple classifier systems provide an effective way to improve pattern recognition performance. In this paper, we use multiple classifier combination to improve LDA for high dimensional data classification. When dealing with the high dimensional data, LDA often suffers from the small sample size problem and the constructed classifier is biased and unstable. Although some approaches, such as PCA+LDA and Null Space LDA, have been proposed to address this problem, they are all at cost of discarding some useful discriminative information. We propose an approach to generate multiple Principal Space LDA and Null Space LDA classifiers by random sampling on the feature vector and training set. The two kinds of complementary classifiers are integrated to preserve all the discriminative information in the feature space.