Selecting Salient Features in High Feature to Exemplar Ratio Conditions

Abstract : We present an approach for identifying salient input features in high feature to exemplar ratio conditions. Basically we modify the SNR saliency-screening algorithm to improve the solution of the optimal salient feature subset problem. We propose that applying the SNR method to randomly selected subsets (SRSS) has a superior potential to identify the salient features than the traditional SNR algorithm has. Two experimental studies are provided to demonstrate the consistency of the SRSS. In the first experiment we used a noise-corrupted version of the Fisher s Iris classification problem. The first experiment designed to prove the fidelity of the SRSS method. The second application is a real-life industrial problem. The salient features of this dataset are not known beforehand. We compared the performances of the salient feature subsets created by the traditional SNR and the SRSS method. We also realized that the SRSS algorithm improved the current solution to this industrial application.