Disentangled Negative Sampling for Collaborative Filtering

Adaptive Hardness Negative Sampling for Collaborative Filtering

“…1. Specifically, for a positive pair (u, i + ), we follow conventional methods (Chen et al 2022;Lai et al 2023) and adapt the two-pass sampling idea, which first randomly samples a fixed size of uninteracted items to form a candidate set, and then selects a negative sample from the candidate set according to predefined rating functions and sampling rules.…”

“…Hard negative sampling methods emphasize the importance of oversampling hard negative samples to speed up the training process and find more precise delineations of user interests. More specifically, it is achieved by either assigning higher sampling probabilities to items with larger predicted scores (Zhang et al 2013;Ding et al 2020;Huang et al 2021;Zhu et al 2022;Lai et al 2023;Shi et al 2023;Zhao et al 2023) or leveraging adversarial learning techniques (Wang et al 2017;Cai and Wang 2018;Park and Chang 2019). For instance, dynamic negative sampling (DNS) (Zhang et al 2013) selects the item with the highest predicted score in a candidate negative sample set.…”

“…SRNS (Ding et al 2020) oversamples items with both high predicted scores and high variances to tackle the false negative problem. DENS (Lai et al 2023) disentangles relevant and irrelevant factors of items and identifies the best negative samples with a factor-aware sampling strategy. Instead of directly selecting negative samples from uninteracted items, MixGCF (Huang et al 2021) synthesizes hard negative samples by mixing positive information into negative samples, which further improves the performance.…”

“…In many cases, it is not always possible to obtain a large amount of high-quality explicit feedback. As a result, implicit feedback, such as clicks or purchases, has become a default choice to train a CF model (Lai et al 2023). In implicit feedback, each observed interaction normally indicates a user's interest in an item and corresponds to a positive training sample.…”

“…For example, random negative sampling (RNS) (Rendle et al 2009) chooses uninteracted items with equal probability, and popularity-biased negative sampling (PNS) (Chen et al 2017;Wu et al 2019) adopts itempopularity-biased distributions to favor popular items. The other line is hard negative sampling, such as dynamic negative sampling (DNS) (Zhang et al 2013) and disentangled negative sampling (DENS) (Lai et al 2023), which focuses on selecting hard negative samples that are difficult to be distinguished from the positive samples with dynamic distributions. Such hard negative samples can provide more informative training signals so that user interests can be better characterized (Xu et al 2022).…”