Adversarial Propensity Weighting for Debiasing in Collaborative Filtering

Kuiyu Zhu, Tao Qin, Pinghui Wang, Xin WangIJCAI 2025 · DOI

Introduction

Modern recommender systems primarily rely on observational data, such as user click-through logs or purchase histories. However, this data is frequently skewed by two pervasive biases: selection bias and popularity bias. Selection bias occurs because users only interact with a fraction of available items, making unobserved interactions ambiguous. Popularity bias leads models to favor frequently interacted items, creating a feedback loop that marginalizes niche content.

Existing debiasing techniques, such as Inverse Propensity Scoring (IPS), often struggle with high variance and training instability, particularly when dealing with rare items. Furthermore, many methods fail to disentangle a user's genuine interest from bias-driven signals. To address these challenges, this paper introduces APWCF (Adversarial Propensity Weighting for Collaborative Filtering), a framework designed to jointly mitigate selection and popularity biases through dynamic weighting and adversarial learning.

Methods

The APWCF framework integrates two primary components into standard collaborative filtering backbones, such as Matrix Factorization or LightGCN.

The APWCF framework architecture featuring the Dynamic Propensity Factor (DPF) and Bias Adversarial Learning (BAL) modules.

Dynamic Propensity Factor (DPF)

Traditional IPS relies on static popularity counts, which can lead to extreme weights and numerical instability. APWCF introduces the Dynamic Propensity Factor (DPF), which learns propensity scores dynamically from user and item embeddings. By incorporating a clipping mechanism, DPF prevents extreme values, ensuring stable training while appropriately weighting long-tail items to counteract selection bias.

Bias Adversarial Learning (BAL)

To address popularity bias at the representation level, the Bias Adversarial Learning (BAL) module employs an adversarial training strategy. A bias discriminator attempts to identify an item's popularity based on its embedding. Simultaneously, a Gradient Reversal Layer (GRL) encourages the encoder to produce embeddings that are uninformative to the discriminator regarding popularity. This process effectively "scrubs" popularity signals, forcing the model to rely on intrinsic user-item compatibility.

Evaluation

The researchers evaluated APWCF across five diverse real-world datasets: ML-100K, Yahoo!R3, KuaiRec, Douban-book, and Yelp2018. These benchmarks cover various domains, including movies, music, short videos, and books, providing a comprehensive assessment of the model's generalizability.

Key Insights and Ideas

The success of APWCF stems from several critical design choices:

  • Stability through Dynamic Weighting: By making propensity estimation dynamic and clipped, the framework transforms a volatile weighting process into a stable optimization task.
  • Bias as Noise: BAL treats bias not just as a weighting problem but as a signal to be removed from internal representations. This ensures that the learned embeddings reflect "pure" user interest.
  • Synergistic Effects: Ablation studies confirm that DPF and BAL address different facets of bias. DPF manages interaction frequency (selection bias), while BAL refines item representations (popularity bias).
Ablation study results on Yahoo!R3 and KuaiRec datasets showing the impact of removing DPF and BAL modules.

Key Results

APWCF demonstrated superior performance, outperforming ten major baselines—including IPS, MACR, and PDA—in 34 out of 40 measured metrics.

  • Performance Gains: Compared to the state-of-the-art PDA method, APWCF achieved improvements of up to 5.42% in Recall@10 and 8.60% in NDCG@10.
  • Robustness: The model showed exceptional stability in "high-intervention" scenarios, where test distributions significantly differ from biased training data. This suggests that APWCF successfully learns underlying preferences rather than merely memorizing popular trends.

Limitations and Future Work

While effective, APWCF has certain limitations. Its performance is slightly reduced on very small datasets when using shallow models, indicating that the adversarial components require sufficient data to converge. Additionally, the framework currently focuses on static debiasing; future research could explore its application to dynamic feedback loops where system recommendations continuously influence future training data. Finally, the added complexity of adversarial layers introduces computational overhead, though the significant accuracy gains often justify this trade-off.

Summary updated

BibTeX
@inproceedings{zhu2025adversarial,
  author = {Kuiyu Zhu and Tao Qin and Pinghui Wang and Xin Wang},
  title = {{Adversarial Propensity Weighting for Debiasing in Collaborative Filtering}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/412},
}
Figure