Structure Aware Incremental Learning with Personalized Imitation Weights for Recommender Systems

Recommender systems now consume large-scale data and play a significant role in improving user experience. Graph Neural Networks (GNNs) have emerged as one of the most effective recommender system models because they model rich relational information. The ever-growing volume can make training GNNs prohibitively expensive. To address this, previous attempts propose to train GNN incrementally new blocks arrive. Feature structure knowledge distillation techniques been explored allow fast incremental fashion while alleviating catastrophic forgetting problem. However, preserving same amount historical information for all users is sub-optimal since it fails take into account dynamics each user's change preferences. For whose interests shift substantially, retaining too much old overly constrain model, preventing from quickly adapting users’ novel interests. In contrast, who static preferences, performance benefit greatly long-term preferences possible. this work, we strategy that adaptively learns personalized imitation weights balance contribution recent be distilled time periods. We demonstrate effectiveness learning via comparison on five diverse datasets three state-of-art based systems. shows consistent improvement over competitive techniques.