Revisiting Negative Sampling vs. Non-sampling in Implicit Recommendation

Recommendation systems play an important role in alleviating the information overload issue. Generally, a recommendation model is trained to discern between positive (liked) and negative (disliked) instances for each user. However, under open-world assumption, there are only but no from users’ implicit feedback, which poses imbalanced learning challenge of lacking samples. To address this, two types strategies have been proposed before, sampling strategy non-sampling strategy. The first samples missing data (i.e., unlabeled data), while regards all as negative. Although known be essential algorithm performance, in-depth comparison has not sufficiently explored by far. bridge this gap, we systematically analyze work. Specifically, theoretically revisit objection non-sampling. Then, with careful setup various representative methods, explore performance different scenarios. Our results empirically show that although widely applied recent models, it non-trivial uniform methods comparable methods. Finally, discuss scalability complexity present some open problems future research topics worth being further explored.