Fitness Functions for Forward Selection: Application in Seven Data Sets

Abstract

Feature selection (FS) is an important part of quantitative structure-activity relationship (QSAR) research. Fitness function is an important factor in FS. Adjusted r2 (r2adj) and Akaike Information Criterion (AIC) are two commonly used fitness functions. Four fitness function, RIC1, RIC2, RIC3 and RIC4 based on the maximization of a ratio of r2adj and AIC are proposed here. A Forward Selection method based on the fitness function was applied to QSAR modelling study of 7 datasets with more than ten thousand samples in total, which was compared with the Forward Selection method based on three other fitness functions (r2adj, AIC and BIC). Final multilinear models were obtained, and 16 performance tests were carried out. Among the 16 performance tests of all the models, the RIC2 model had 3 indexes, which are the best, and there were no worst indicators. The results show that the RIC2 model is suitable for prediction.