<ul><li>REMEDI (Relative Feature Enhanced Meta-Learning with Distillation for Imbalanced Prediction) is a novel multi-stage framework designed to predict future vehicle purchases among existing owners despite extreme class imbalance and complex behavioral patterns.</li><li>REMEDI first trains diverse base models to capture different aspects of user behavior, then introduces relative performance meta-features for effective model fusion through a hybrid-expert architecture.</li><li>Finally, REMEDI distills the ensemble's knowledge into a single efficient model through supervised fine-tuning with MSE loss, outperforming baseline approaches and achieving the business target of identifying around 50% of actual buyers within the top 60,000 recommendations at about 10% precision.</li><li>Evaluated on approximately 800,000 vehicle owners, REMEDI demonstrates effectiveness for imbalanced prediction in industry settings by significantly improving prediction accuracy and maintaining deployment efficiency.</li></ul>

REMEDI: Relative Feature Enhanced Meta-Learning with Distillation for Imbalanced Prediction

Discover more