• Learns user-personalized imitation weights that preserve historical preferences during incremental updates.
• Boosts recall and NDCG scores on four benchmark datasets with limited compute overhead.