Functional gene prediction with vital reduced features: Further topics for feature reduction and evaluation criteria for classifiers

Abstract

Aiming at the prediction of protein solubility, four feature reduction methods are discussed in this paper, including Correlation coefficient method, Filter method, Relief method and Genetic method. With the Top 100 features discovered by genetic method, the best classifier achieves the accuracy of 86% and MCC of 0.7236 in Jackknife test. Moreover, further discussions about feature reduction and classifier reliability evaluation criteria are given. The author claim the exclusive importance of capacity of expansion prediction for classifiers.