CIPL: Counterfactual Interactive Policy Learning to Eliminate Popularity Bias for Online Recommendation

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Popularity bias is a persistent issue in recommender systems (RSs), particularly in online interactive settings where bias amplification is exacerbated by feedback loops.
• Existing research primarily addresses offline recommendation and often overlooks the temporal dependencies crucial for online interactive systems.

Purpose of the Study:

• To propose a novel method, counterfactual interactive policy learning (CIPL), for eliminating popularity bias in online recommendation scenarios.
• To address the limitations of static causal analysis by incorporating temporal dependencies in interactive recommendation systems.

Main Methods:

• Formulation of a novel temporal causal graph (TCG) to model causal relationships and guide counterfactual inference within interactive recommender models.
• Development of the CIPL method using an actor-critic framework and an online interactive environment simulator for training.
• Estimation of item popularity's causal effect on prediction scores at each interaction step and removal of popularity bias during the testing phase.

Main Results:

• Extensive experiments on three public benchmarks demonstrate the effectiveness of the proposed CIPL method.
• The CIPL method achieves state-of-the-art performance in eliminating popularity bias in online interactive recommendation systems.
• The temporal causal graph (TCG) successfully captures and accounts for temporal dependencies, improving bias mitigation.

Conclusions:

• The proposed CIPL method offers a robust solution for addressing popularity bias in dynamic online recommendation environments.
• Incorporating temporal causal graphs is essential for accurately modeling and mitigating bias in interactive systems.
• This research advances the field of fair and effective recommender systems by tackling a critical challenge in real-world applications.