New Framework of Location Model for Mixed Variables Classification in the Presence of Outliers

Abstract

Classification that contains a mix of binary and continuous variables is a common challenge in statistical modelling. The Location Model (LM) was developed to address this challenge by discriminating groups based on continuous variables classified by multinomial cells derived from binary variables. However, a key limitation of the LM is its susceptibility to outliers, which can significantly degrade its classification accuracy, leading to high misclassification rates. To address this issue, this study aims to develop a robust framework for classifying mixed variables in the presence of outliers, with a focus on improving the LM. The developed framework introduces the Robust Location Model (RLM_med) based on the median, which enhances the model's resistance to outliers by employing a median estimator combined with a robust covariance matrix. The RLM_med framework is anticipated to outperform existing models, offering superior classification performance in datasets that contain mixed variables and are affected by outliers.