New Research Introduces Multiclass Local Calibration Framework Using Jensen-Shannon Distance

A new research paper introduces multiclass local calibration to address proximity bias in machine learning models. The work establishes formal definitions connecting local calibration with the stringent concept of strong calibration. Researchers identify vulnerabilities in existing evaluation metrics when applied to multiclass local calibration scenarios. The paper proposes using Jensen-Shannon distance as a methodological approach to improve calibration across all predicted probabilities. This addresses the critical shortcoming where predictions in sparse feature space regions become systematically miscalibrated. The research focuses on developing trustworthy ML models whose predicted probabilities accurately reflect true-class frequencies. The work analyzes theoretical pitfalls in current multiclass calibration approaches that lack distance considerations among inputs. The paper is available as arXiv:2510.26566v2 with announcement type replace-cross.