The study addresses the limitations of current temporal graph models, which often involve only the most recent neighbors, leading to incomplete and biased data. The authors propose a novel framework called Recurrent Temporal Revision (RTR) for temporal neighbor aggregation. It uses a recurrent neural network with node-wise hidden states to incorporate information from all historical neighbors. This approach has shown superior theoretical expressiveness and has improved performance in real-world applications, notably achieving a +9.6% improvement in averaged precision in a real-world Ecommerce dataset over existing methods on 2-layer models.
Publication date: 22 Sep 2023
Project Page: https://arxiv.org/abs/2309.12694v1
Paper: https://arxiv.org/pdf/2309.12694
