Component and feature selection in finite mixture regression models: a self-paced learning approach
Abstract
Component and feature selection are essential when considering finite mixture regression. We introduce a penalised self-paced learning approach for order and feature selection. Self-paced learning uses weighted likelihood to evaluate model fitting, with low contributions from non-typical observations. It starts with a low tolerance for non-typical observations, increasing the tolerance during later stages of the fitting process. This approach yields more robust estimates. Adding appropriate penalty terms to the weighted likelihood function facilitates order and feature selection. We propose an EM-type algorithm to maximise the penalised self-paced learning weighted likelihood function, wrapped within the self-paced learning algorithm, mitigating the impact of non-typical observations. The properties of the proposed estimation procedure, including robustness properties, are evaluated using an extensive simulation study. The proposed estimation procedure is also demonstrated using real data.