Publications

Snapshot-based dynamic GNNs are typically temporal-first or spatial-first: the rigid sequencing forces the second stage to consume an already-compressed summary from the first, so the message-passing operator never gets to weigh a neighbour's contribution by that neighbour's past trajectory. SiST-GNN instead fuses both signals inside a single message-passing operation — each node carries a recurrent hidden state summarising its history, which is paired with its current feature vector as two nodes joined by a cross-time edge, and a standard graph convolution over this temporally augmented graph yields the updated representation. Against fourteen link-prediction baselines, SiST-GNN improves on the strongest prior method by 1–18% in the fixed-split setting and leads on five of six datasets in the live-update setting (1–158%); on dynamic node classification it beats the best discrete-time baseline by 7–23% while matching continuous-time methods that consume raw event streams.

This paper addresses the computational efficiency challenges in Graph Neural Networks by proposing a novel graph coarsening technique. Our approach significantly reduces inference time and memory while maintaining model performance, making GNNs more practical for real-world applications with large-scale graphs.

We propose a novel approach to enhance the expressivity of Graph Neural Networks beyond the traditional Weisfeiler-Leman hierarchy limitations. Our localization-based method improves the ability of GNNs to distinguish between different graph structures while maintaining computational efficiency.