Unbiased Recommender Learning for Enhanced Relevance and Coverage

In recent years, the effect of cognitive bias on recommender systems has become a major concern. These biases are commonly found in data collected through the feedback loop. Ignoring them may cause problems such as homogenization among users and increasing disparity among items. Therefore, it is necessary to address these biases for sustainable and effective operation of the recommender system. To address exposure bias, a method has been proposed that maximizes the relevance of recommendations by integrating matrix factorization (MF) with an inverse propensity score (IPS) estimator in the pointwise loss function. Although MF achieves a high recommendation accuracy, it tends to bias interactions, for instance, by limiting exposure opportunities for certain items. In this study, we demonstrate that incorporating factorization machines (FM), which can handle various features, into the IPS estimator can improve both the relevance and coverage of recommendations—attributes that typically have a trade-off relationship—using biased data. Experimental results from semi-synthetic data indicate that our approach can significantly enhance item coverage while maintaining the high relevance of recommendations, particularly in a large-scale data environment. This advancement allows many users to discover relevant content from biased data and increases exposure opportunities for a broader range of content.