Data Analysis using Independent Component Analysis

Abstract

In this work, a large data set has been considered for computational modeling using Independent Component Analysis (ICA) and data mining techniques. A technique using negentropy maximization has been proposed for Blind Source Separation (BSS). It has been successfully used for separating the image and audio signals. Simulations have been done to describe a technique for measuring information gain from large data sets using entropy analysis and information theoretic measures. Due to the increasing requirements in the commercial sectors, the efficient data mining and knowledge discovery techniques bear a very important role. The feature extraction task is done to select most of the important information that represents the original data. The neural techniques make it possible to extract the optimal and important features for future processing and decision-making. Statistical techniques such as Principal Component Analysis (PCA) and ICA may be used to analyze high dimensional data. Optimizing the objective functions like mutual information, joint entropy, negentropy, kurtosis etc can derive the iterative algorithms for ICA. We have proposed a new approach for ICA based on information theory for processing large data sets. The technique has been implemented to extract features from multidimensional databases. The approach is based entirely on measured entropies of the system and minimization of mutual information. It has been observed that the proposed technique yields the learning rule for fixed non-linearity. newline... newline