Parallelizing ICA for Text-Feature Extraction in PC Clusters

Independent component analysis (ICA) is widely applied to the blind source separation problems. ICA is a method for finding hidden factors or components from linearly or non-linearly mixed data variables. It looks for components that are at the lowest dependences of each other, that is, statistically independent or nongaussian. However, the estimation of independent components is a time consuming process. We present a parallel ICA algorithm in this paper which conducts the estimation of the weight matrix in an MPI environment consisting of 2 and 4 processors. The parallel ICA algorithm we developed is based on the FastICA algorithm which is the most efficient ICA algorithm so far. We have used ICA in text feature extraction and the time complexity increased with the increasing number of features. The Results show that the parallel ICA could speedup the computation of independent components and the classification results using the features extracted by it has 90% match with the categories classified manually. The parallel ICA algorithm had distributed the computation load to multiple processor without losing too much of the accuracy.