Generalized Negative Sampling for Implicit Feedback in Recommendation

In a typical model-based collaborative filtering with implicit feedback, negative sampling is getting more and more popular to obtain negative labeled inputs from massive unobserved data. However, this approach wrongly samples false negatives, resulting in unacceptable recommender model. In this work, we first identify a situation where false negatives are problematic and estimate their impact on a model accuracy. Then, we take our negative sampling as a classification task and demonstrate that a recommender model and a negative sampling method actually share the same goal: identification of false negatives from other unobserved data. We also estimate the actual upper bound of the accuracy improvements with a feasible negative sampling. Lastly, we propose a generalized negative sampling that can alleviate the impact of false negatives by introducing two robustness against false negatives: self-sampling and dynamic sub-sampling. In user-item interaction matrix, experimental results on publicly available datasets show that our approach outperforms some state-of-the-arts with statistical significance.