A Responsible AI approach for designing resilient classifier to handle incomplete data

Abstract

Missing values can greatly affect analyses and decision-making in many fields. In the context of Responsible Artificial Intelligence (AI), ensuring the robustness of machine learning models is essential because Responsible AI emphasizes reliability and interpretability in decision-making processes. However, traditional imputation and ensemble learning methods often fail to preserve critical relationships between independent and dependent variables, introducing bias or noise into the data and undermining the development of robust classification models. To address these challenges, we propose a novel classification approach that aligns with Responsible AI principles. Our Resilient Decision Tree classifier is specifically designed to handle incomplete datasets. We employ subspace classifiers that operate on different non overlapping subsets of features without relying on imputation. By combining these subspace models into a weighted ensemble classifier, we enhance prediction accuracy for test datasets with missing values. The experimental results obtained on real-life and synthetic datasets demonstrate that our methodology produces an effective ensemble classifier.