Simplify and Robustify Negative Sampling for Implicit Collaborative Filtering

Since implicit feedback data contain positive instances only, the implicit CF problem is also related to learning from positive-unlabeled (PU) data. The first three authors have equal contributions. However, the size of unlabeled data in implicit CF can approach to nearly a product of user count and item count, making above non-sampling approach become unaffordable in terms of learning efficiency. Negative sampling approaches have also been widely adopted in other domains of embedding learning for text, graph, etc. Motivated by these works that tend to leverage a simple model for Our SRNS's hyper-parameters can be divided into three parts: (1) sampling related part, including In synthetic noise experiments, since we do not explicitly split a validation set on synthetic data, we draw two different train/test splits. In real data experiments, we conduct the standard procedure to split train/validation/test set.