Extraction of independent discriminant features for data with asymmetric distribution

Standard unsupervised linear feature extraction methods find orthonormal (PCA) or statistically independent (ICA) latent variables that are good for data representation. These representative features may not be optimal for the classification tasks, thus requiring a search of linear projections that can give a good discriminative model. A semi-supervised linear feature extraction method, namely dICA, had recently been proposed which jointly maximizes the Fisher linear discriminant (FLD) and negentropy of the extracted features [Dhir and Lee in Discriminant independent component analysis. In: Proceedings of the international conference intelligent data engineering and automated learning, LNCS 5788:219-225 (Full paper is submitted to IEEE Trans. NN) 2009]. Motivated by the independence and unit covariance of the extracted dICA features, maximizing the determinant of between-class scatter of the features matrix is theoretically the same as the maximization of FLD. This also reduces the computational complexity of the algorithm. In this paper, we concentrate on text databases that follow inherent exponential distribution. Approximation and the maximization of negentropy for data with asymmetric distribution is discussed. Experiments on the text categorization problem show improvements in classification performance and data reconstruction.