Go to ICLR 2025 Conference homepageRisk Aware Negative Sampling in Link Prediction
It is commonly believed that Message Passing Neural Networks (MPNNs) struggle in link prediction settings due to limitations in their expressive power. Recent work has focused on developing more expressive model classes, which are capable of learning link representations through techniques such as labeling tricks, the inclusion of structural features, or the use of subgraph methods. These approaches have yielded significant performance improvements across a range of benchmark datasets. However, an interesting question remains: have we fully wrung out the performance by optimizing the other aspects of the training process? In this work, we present results that indicate that significant amounts of model performance have been left on the table by the use of easy negative-samples during training. We theoretically explore the generalization gap and excess risk to quantify the performance loss caused by easy negatives. Motivated by this analysis, we introduce Risk Aware Negative Sampling in Link Prediction (RANS), which efficiently performs dynamic hard-negative-mining. Empirical results show that a simple GCN augmented by RANS realizes between 20% and 50% improvements in predictive accuracy when compared with the same model trained with standard negative samples.