Interpolative Distillation for Unifying Biased and Debiased Recommendation

Most recommender systems evaluate model performance offline through either: 1) normal biased test on factual interactions; or 2) debiased test with records from the randomized controlled trial. In fact, both tests only reflect part of the whole picture: factual interactions are collected from the recommendation policy, fitting them better implies benefiting the platform with higher click or conversion rate; in contrast, debiased test eliminates system-induced biases and thus is more reflective of user true preference. Nevertheless, we find that existing models exhibit trade-off on the two tests, and there lacks methods that perform well on both tests.

In this work, we aim to develop a win-win recommendation method that is strong on both tests. It is non-trivial, since it requires to learn a model that can make accurate prediction in both factual environment (i.e., normal biased test) and counterfactual environment (i.e., debiased test). Towards the goal, we perform environment-aware recommendation modeling by considering both environments. In particular, we propose an Interpolative Distillation (InterD) framework, which interpolates the biased and debiased models at user-item pair level by distilling a student model. We conduct experiments on three real-world datasets with both tests. Empirical results justify the rationality and effectiveness of InterD, which stands out on both tests especially demonstrates remarkable gains on less popular items.