Enhanced Bayesian Personalized Ranking for Robust Hard Negative Sampling in Recommender Systems

TL;DR

This paper introduces an enhanced Bayesian Personalized Ranking objective, named as Hard-BPR, which is specifically crafted for dynamic hard negative sampling to mitigate the influence of false negatives, and is simple yet efficient for real-world deployment.

Abstract

In implicit collaborative filtering, hard negative mining techniques are developed to accelerate and enhance the recommendation model learning. However, the inadvertent selection of false negatives remains a major concern in hard negative sampling, as these false negatives can provide incorrect information and mislead the model learning. To date, only a small number of studies have been committed to solve the false negative problem, primarily focusing on designing sophisticated sampling algorithms to filter false negatives. In contrast, this paper shifts its focus to refining the loss function. We find that the original Bayesian Personalized Ranking (BPR), initially designed for uniform negative sampling, is inadequate in adapting to hard sampling scenarios. Hence, we introduce an enhanced Bayesian Personalized Ranking objective, named as Hard-BPR, which is specifically crafted for dynamic hard negative sampling to mitigate the influence of false negatives. This method is simple yet efficient for real-world deployment. Extensive experiments conducted on three real-world datasets demonstrate the effectiveness and robustness of our approach, along with the enhanced ability to distinguish false negatives.

Figures (4)

• Figure 1: (a) The function curves of $\sigma(\cdot)$ and $g(\cdot)$. (b) The function curves of $\Delta_\sigma(\cdot)$ and $\Delta_g(\cdot)$.
• Figure 2: The training curves of RNS, DNS and the proposed method on the test sets of three datasets. (a)(b)(c) With MF as the base model. (d)(e)(f) With LightGCN as the base model.
• Figure 3: The distributions of predicted true negative scores and predicted false negative scores, employing different base recommendation models and different negative sampling methods (RNS, DNS and the proposed method).
• Figure 4: Parameter study of the proposed method. Experiments on Tmall under MF. The left column: curves of $\Delta_g(\cdot)/c$. The right column: training curves of our method.