UFNRec: Utilizing False Negative Samples for Sequential Recommendation
Xiao-Yang Liu, Chong Liu, Pinzheng Wang, Rongqin Zheng, Lixin Zhang, Leyu Lin, Zhijun Chen, Liangliang Fu · Society for Industrial and Applied Mathematics eBooks · 2023
Sequential recommendation models are primarily optimized to distinguish positive samples from negative ones during training. Thus, negative instances sampled from enormous unlabeled data are essential in learning the evolving user preferences through historical records. Except for randomly sampling negative samples from a uniformly distributed subset, many delicate methods have been proposed to mine negative samples with high quality. However, due to the inherent randomness of negative sampling, false negatives are inevitably collected in model training. Current strategies mainly focus on removing such false negatives, which leads to overlooking potential user interests, lack of recommendation diversity, less model robustness, and suffering from exposure bias. To this end, we propose a novel method that can Utilize False Negative samples for sequential Recommendation (UFNRec), which thoroughly explores the leverage of false negatives. We first devise a simple strategy to extract false negatives from true negatives and directly reverse the labels of false negatives. To avoid extra noise from reversed samples, we restrict false negatives in the output space by an EMA operation and a consistency regularization loss. To the best of our knowledge, this is the first work to utilize false negatives instead of simply removing them for sequential recommendation. Both offline and online experiment results demonstrate that UFNRec can effectively draw information from false negatives and further improve the performance of SOTA models. Recently, we have deployed UFNRec on real-world recommendation servings. The code is available at https://github.com/UFNRec-code/UFNRec.