Machine-learning evaluation demonstrates balanced trade-offs across predictive accuracy and diverse fairness metrics in motor insurance, suggesting viable multi-objective pricing models.
Key Points
Develop and evaluate a multi-objective optimization framework that simultaneously balances predictive accuracy with group, individual, and counterfactual fairness metrics in insurance risk pricing.
Constructed a multi-objective framework combining fairness-aware base models and approximated a four-objective Pareto front using the Non-dominated Sorting Genetic Algorithm II (NSGA-II).
Applied the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) to identify a compromise solution along the Pareto front.
Trained and evaluated generalized linear models (GLMs) against Extreme Gradient Boosting (XGBoost) and the multi-objective ensemble across two motor insurance datasets.
Extreme Gradient Boosting (XGBoost) enhanced predictive accuracy relative to standard generalized linear models but exacerbated disparities across several fairness metrics.
The proposed ensemble reached a balanced compromise among the four objectives, successfully improving aggregate fairness without causing severe degradation in predictive accuracy.