BPL: Bias-adaptive Preference Distillation Learning for Recommender System

TL;DR

BPL addresses the bias-induced gap between true user preferences and observed explicit feedback by formulating a risk-bounded learning objective with three terms and introducing dual distillation strategies. It combines reliability-filtered self-distillation for unrated data with confidence-penalized distillation from a biased teacher, adaptively weighted by $\mathcal{S}^1$-affinity, to optimize both factual and counterfactual test performance. Empirical results on three real-world datasets show that BPL achieves superior balance across factual and counterfactual environments, with ablations confirming the necessity of each component and the penalty term for preventing overfitting to biased predictions. The method offers practical impact for recommender systems by improving generalization across diverse data regimes without requiring unbiased training data, and it provides robust guidance on hyperparameter choices and affinity estimation. Overall, BPL advances debiasing-aware preference learning by jointly optimizing for accuracy on actual user interactions and unbiased evaluation scenarios.

Abstract

Recommender systems suffer from biases that cause the collected feedback to incompletely reveal user preference. While debiasing learning has been extensively studied, they mostly focused on the specialized (called counterfactual) test environment simulated by random exposure of items, significantly degrading accuracy in the typical (called factual) test environment based on actual user-item interactions. In fact, each test environment highlights the benefit of a different aspect: the counterfactual test emphasizes user satisfaction in the long-terms, while the factual test focuses on predicting subsequent user behaviors on platforms. Therefore, it is desirable to have a model that performs well on both tests rather than only one. In this work, we introduce a new learning framework, called Bias-adaptive Preference distillation Learning (BPL), to gradually uncover user preferences with dual distillation strategies. These distillation strategies are designed to drive high performance in both factual and counterfactual test environments. Employing a specialized form of teacher-student distillation from a biased model, BPL retains accurate preference knowledge aligned with the collected feedback, leading to high performance in the factual test. Furthermore, through self-distillation with reliability filtering, BPL iteratively refines its knowledge throughout the training process. This enables the model to produce more accurate predictions across a broader range of user-item combinations, thereby improving performance in the counterfactual test. Comprehensive experiments validate the effectiveness of BPL in both factual and counterfactual tests. Our implementation is accessible via: https://github.com/SeongKu-Kang/BPL.

Figures (10)

• Figure 1: Mean squared errors of various preference learning methods on factual and counterfactual test across three real-world datasets. We compare four different methods: (a) Standard training without debiasing learning, (b) Stable-MRDR STABLEDR, a recent debiasing method, (c) InterD InterD, a representative knowledge distillation method, and (Purple) BPL, the proposed method.
• Figure 2: Relationships between $\mathcal{S}^1$-affinity (i.e., $\hat{p}(s_{ui}=1)$) and model prediction errors with varying learning strategies.
• Figure 3: Relationships between $\mathcal{S}^1$-affinity (i.e., $\hat{p}(s_{ui}=1)$) and three factors that affect rating process.
• Figure 4: Illustration of BPL. (a) The overall framework, (b) Dual distillation learning with unrated data. Best viewed in color.
• Figure 5: Comparison of filtering effects. (left) post-filtering error ratio, which denotes the ratio of average errors in the filtered dataset to the average errors in the total dataset, (right) recommendation performance. Results on Yahoo!R3-RCT.