BPL: Bias-adaptive Preference Distillation Learning for Recommender System
On the map: Knowledge distillation → Distillation for recommender systems
Abstract (arXiv)
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.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Distillation for recommender systems, one of 12 directions in Knowledge distillation. Builds on 34 of its 61 references that are in the map. First posted on arXiv 17 Oct 2025.
Builds on
34 of 61 references.
Closest ancestors · 2
Same direction, newest first
- 2023Distillation from Heterogeneous Models for Top-K Recommendation2 Mar
- 2022Unbiased Knowledge Distillation for Recommendation27 Nov
Other references · 32
- 2015Distilling the Knowledge in a Neural Networkcited by 10,324
- 2015Domain-Adversarial Training of Neural Networkscited by 3,044
- 2017Neural Collaborative Filteringcited by 1,111
- 2017Regularizing Neural Networks by Penalizing Confident Output Distributionscited by 464
- 2020Towards Understanding Ensemble, Knowledge Distillation and Self-Distillation in Deep Learningcited by 297
- 2016Recommendations as Treatments: Debiasing Learning and Evaluationcited by 248
- 2020Bias and Debias in Recommender System: A Survey and Future Directionscited by 240
- 2022FreeMatch: Self-adaptive Thresholding for Semi-supervised Learningcited by 232
- 2020Classes Matter: A Fine-grained Adversarial Approach to Cross-domain Semantic Segmentationcited by 142
- 2022KuaiRec: A Fully-observed Dataset and Insights for Evaluating Recommender Systemscited by 96
- 2015Modeling User Exposure in Recommendationcited by 69
- 2021AutoDebias: Learning to Debias for Recommendationcited by 68
20 more
- 2020ESAM: Discriminative Domain Adaptation with Non-Displayed Items to Improve Long-Tail Performancecited by 30
- 2019Asymmetric Tri-training for Debiasing Missing-Not-At-Random Explicit Feedbackcited by 28
- 2020DE-RRD: A Knowledge Distillation Framework for Recommender Systemcited by 28
- 2021Enhanced Doubly Robust Learning for Debiasing Post-click Conversion Rate Estimationcited by 28
- 2022StableDR: Stabilized Doubly Robust Learning for Recommendation on Data Missing Not at Randomcited by 24
- 2021Popularity Bias Is Not Always Evil: Disentangling Benign and Harmful Bias for Recommendationcited by 23
- 2023Invariant Collaborative Filtering to Popularity Distribution Shiftcited by 23
- 2022TDR-CL: Targeted Doubly Robust Collaborative Learning for Debiased Recommendationscited by 18
- 2023Balancing Unobserved Confounding with a Few Unbiased Ratings in Debiased Recommendationscited by 18
- 2022UKD: Debiasing Conversion Rate Estimation via Uncertainty-regularized Knowledge Distillationcited by 14
- 2021Re-energizing Domain Discriminator with Sample Relabeling for Adversarial Domain Adaptationcited by 12
- 2021GLocal-K: Global and Local Kernels for Recommender Systemscited by 11
- 2022Undoing the Damage of Label Shift for Cross-domain Semantic Segmentationcited by 11
- 2022Multiple Robust Learning for Recommendationcited by 11
- 2022Mitigating Hidden Confounding Effects for Causal Recommendationcited by 8
- 2024A Teacher-Free Graph Knowledge Distillation Framework with Dual Self-Distillationcited by 8
- 2024Debiased Recommendation with Noisy Feedbackcited by 8
- 2024Doubly Calibrated Estimator for Recommendation on Data Missing Not At Randomcited by 7
- 2019Towards Resolving Propensity Contradiction in Offline Recommender Learningcited by 6
- 2023CDR: Conservative Doubly Robust Learning for Debiased Recommendationcited by 2
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2024Preference Diffusion for Recommendation
- 2022Bilateral Self-unbiased Learning from Biased Implicit Feedback
- 2025What Makes LLMs Effective Sequential Recommenders? A Study on Preference Intensity and Temporal Context
- 2026P3Rec: Distilling Prior--Posterior Preference Reasoning for LLM-based Recommendation
- 2026Consensus vs. Dissent: Dynamic LLM Modeling of Subjective Preferences in Group Recommenders
- 2025Neutralizing Popularity Bias in LLM-based Recommendation via Counterfactual Reasoning Guidelines
- 2022Cross-Task Knowledge Distillation in Multi-Task Recommendation
- 2026Beyond Factual Correctness: Mitigating Preference-Inconsistent Explanations in Explainable Recommendation