Price-aware debiased learning model for recommendation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Popularity bias in recommender systems leads to over-reliance on item popularity, decreasing performance.
• Existing debiasing methods overlook the beneficial aspect of popularity bias, potentially harming recommendations.

Purpose of the Study:

• To investigate the role of item prices in creating a beneficial popularity bias.
• To propose a method for disentangling beneficial price-aware popularity bias from harmful bias.
• To improve recommender system performance by addressing overlooked debiasing aspects.

Main Methods:

• Developed a causal graph and used causal intervention to understand price-aware popularity bias.
• Proposed the Price-Aware Popularity-Debiased Model (PAPDM) incorporating user purchasing capacity and willingness.
• Introduced a popularity-based negative sampling method to identify true negative samples.

Main Results:

• The proposed PAPDM effectively disentangles beneficial price-aware popularity bias.
• PAPDM outperforms recent debiased models on widely used datasets.
• The model successfully reduces harmful popularity bias while retaining beneficial price-related signals.

Conclusions:

• Item prices contribute to a beneficial popularity bias that should be preserved.
• The PAPDM offers a novel approach to debiasing recommender systems by considering price effects.
• Addressing price-aware popularity bias is crucial for enhancing overall recommendation performance.