Correcting for Multifactorial Bias in Recommender Systems: Going Beyond Popularity and Positivity Bias

Recommender systems suffer from multifactorial bias, where user interactions are affected by both item and rating value factors. Existing debiasing methods only consider single-factor biases, which can lead to suboptimal performance when both factors are important. This work proposes a multifactorial bias correction method that estimates propensities based on both item and rating value, and integrates it with an IPS-based optimization approach to provide more robust and effective debiasing.