Efficient and Bias-aware Recommendation with Two-side Relevance for Implicit Feedback

Today’s wide-spread recommendation is usually constructed based on implicit data such as click for easy collection but whether the no clicked data is negative feedback or unobserved positive feedback confuses the model construction. As a response, Relevance Matrix Factorization (Rel-MF) is recently proposed to tackle this problem as well as the missing-not-at-random (MNAR) problem ignored by previous studies. However, Rel-MF meets three problems: limited assumption (LA), negative square loss (NSL) and indiscriminate no click data (INCD). In this paper, we first get rid of Rel-MF’s limited assumption and establish a more general theory by incorporating a defined transformation function which captures the relevance level to our two-side relevance ideal loss, containing Rel-MF’s theory. To resolve the INCD problem and NSL problem, we introduce an adjusting variable and perform normalization, respectively, which is called Naive Solution with Normalization for Rel-MF (NRel-MF). But we then analytically discover that the clipped function proposed by Rel-MF meets the high variance problem. To overcome it, we design a power clipped function and further propose Improved Solution with Power Function for Rel-MF (PRel-MF). Besides, we also explore propensity score estimation from user and hybrid perspectives in contrast to Rel-MF’s sole item perspective. Finally, we also consider and address the computational problem caused by the Rel-MF’s non-sampling strategy. Empirical results verify the effectiveness of our solutions from both performance even in rare items and loss decrease. In broader perspective experiment, decent performance is seen in item perspective with fewer recommended items while in user perspective with more recommended items and hybrid perspective outperforms them in more situations.