Novel model selection criteria for LMARS: MARS designed for biological networks

Abstract

ABSTRACT In higher dimensions, the loop-based multivariate adaptive regression splines (LMARS) model is used to build sparse and complex gene structure nonparametrically by correctly defining its interactions in the network. Also, it prefers to apply the generalized cross-validation (GCV) value as its original model selection criterion in order to select the best model, in turn, represent the true network structure. In this study, we suggest to modify the model selection procedure of LMARS by changing GCV with our Kullback–Leibler information-based criteria, namely, consistent Akaike information criterion (CAIC), CAIC with Fisher information matrix and information complexity. Thereby, we aim to compare the performance of our proposal model selection criteria together with the state-of-art model selection criteria, namely, AIC and Bayesian information criterion by comparing their accuracy with GCV while modelling different dimensional and topological biological networks under both simulated and real datasets.