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

TL;DR

REMEDI is a multi-stage meta-learning framework that effectively predicts rare vehicle purchases by leveraging diverse models, relative performance features, and knowledge distillation, achieving high accuracy in imbalanced industrial scenarios.

Contribution

It introduces a novel multi-stage approach combining ensemble, relative feature fusion, and model distillation specifically for imbalanced prediction tasks.

Findings

Significantly outperforms baseline methods on real-world data.

Achieves 50% recall within top 60,000 recommendations at 10% precision.

Distilled model retains ensemble performance with improved deployment efficiency.

Abstract

Predicting future vehicle purchases among existing owners presents a critical challenge due to extreme class imbalance (<0.5% positive rate) and complex behavioral patterns. We propose REMEDI (Relative feature Enhanced Meta-learning with Distillation for Imbalanced prediction), a novel multi-stage framework addressing these challenges. REMEDI first trains diverse base models to capture complementary aspects of user behavior. Second, inspired by comparative op-timization techniques, we introduce relative performance meta-features (deviation from ensemble mean, rank among peers) for effective model fusion through a hybrid-expert architecture. Third, we distill the ensemble's knowledge into a single efficient model via supervised fine-tuning with MSE loss, enabling practical deployment. Evaluated on approximately 800,000 vehicle owners, REMEDI significantly outperforms baseline approaches, achieving the business target of identifying ~50% of actual buyers within the top 60,000 recommendations at ~10% precision. The distilled model preserves the ensemble's predictive power while maintaining deployment efficiency, demonstrating REMEDI's effectiveness for imbalanced prediction in industry settings.