A0825
Title: Robust finite mixture of regression model selection
Authors: Frans Kanfer - University of Pretoria (South Africa) [presenting]
Andre Kleynhans - University of Pretoria (South Africa)
Sollie Millard - University of Pretoria (South Africa)
Abstract: Finite mixture of regression (FMR) models are widely employed for analysing data observed from heterogeneous subpopulations. Despite the flexibility of FMR models, estimation procedures face challenges, such as pre-specification of the number of components, order selection, and the identification of informative variables and covariates. Standard estimation procedures are also often sensitive to non-typical observations, affecting estimated model performance. An approach is proposed that integrates a penalized information criterion with a self-paced learning (SPL) algorithm. The penalization mechanism enables joint order and variable selection, while the SPL framework regulates the inclusion of observations during the learning process, thereby mitigating the influence of non-typical data. The proposed method is assessed through simulation studies, evaluating its robustness in component number estimation, variable selection, and parameter recovery under varying levels of contamination of non-typical observations. Real-world applications are also considered.
