An Active Learning Based LDA Algorithm for Large-Scale Data Classification

As traditional Linear Discriminant Analysis algorithm runs slowly in large data set, this paper proposed a fast LDA algorithm based on active learning. In the proposed algorithm, the original training set is divided into three parts, i.e. initial training set, correction set and testing set. Secondly, LDA algorithm is running on the initial training set, and the projection vector can be obtained. Thirdly, we select from correction set the samples whose projection is farthest from the mean vector, add them into the initial training set and compute the projection vector again. Repeat this step until the classification precision attains the expected target or the correction set is empty. The simulation experiments on the UCI data set and the MNIST data set show that the proposed algorithm running fast on large data set, and has a good classification precision. © 2016 SERSC.