Generating Comprehensible QSAR Models

This paper presents work in progress from theINFUSIS project and contains initial experimentation, usingpublicly available medicinal chemistry datasets, on obtainingcomprehensible QSAR models. Three techniques are evaluatedon both predictive performance, measured as accuracy, andcomprehensibility, measured as model size. The chosentechniques are J48 decision trees and JRip and Chipper decisionlists. The results show that J48 obtains superior accuracy andthat Chipper performs best of the two decision list algorithms onaccuracy. Furthermore, it is seen that, regarding accuracy, alltechniques benefit from feature reduction, which almost alwaysresults in increased accuracy. Regarding comprehensibility, JRipobtains the smallest models, followed by Chipper, with J48producing the largest models. For model size, feature reduction isnot seen to be universally beneficial; only J48 produces smallermodels for the reduced datasets, while both decision listalgorithms actually produce larger models on average. Theoverall conclusion is that, for these datasets, there exists a definitetradeoff between accuracy and comprehensibility that needs to beinvestigated further