Mitigating Confounding Bias in Practical Recommender Systems With Partially Inaccessible Exposure Status.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Recommender systems (RS) are crucial for user experience on online platforms, learning from user feedback.
• Existing RS methods suffer from confounding bias due to factors influencing item exposure and feedback.
• Previous debiasing strategies struggle to simultaneously capture recommendation-specific and exposure-specific knowledge and handle noisy exposure data.

Purpose of the Study:

• To develop a novel debiasing recommendation approach that addresses limitations in existing methods.
• To effectively capture recommendation-specific, exposure-specific, and common knowledge simultaneously.
• To achieve robustness against partially inaccessible exposure results in recommender systems.

Main Methods:

• Propose a mutual information-based counterfactual learning framework leveraging causal relationships between features, exposure, and ratings.
• Explicitly model relationships among causal factors to capture recommendation-specific, exposure-specific, and common knowledge.
• Employ a pairwise learning strategy for robustness against partially inaccessible exposure data and implement an optimizable loss function.

Main Results:

• The proposed framework successfully captures recommendation-specific, exposure-specific, and common knowledge.
• The method demonstrates robustness towards partially inaccessible exposure results.
• Extensive experiments on public datasets show superior performance in boosting recommendation accuracy.

Conclusions:

• The novel debiasing approach effectively mitigates confounding bias in recommender systems.
• The mutual information-based counterfactual learning framework enhances the ability to capture nuanced user preferences.
• The proposed method offers a robust solution for improving recommendation performance, even with incomplete exposure information.