Publications

Decoding brain states needs models that capture both how neural activity evolves and how groups of brain regions interact. Existing EEG methods largely treat recordings as multivariate time series or reduce functional connectivity to pairwise graphs, leaving dynamic higher-order interactions unmodelled. T-SNN represents an EEG recording as a sequence of evolving simplicial complexes and combines simplicial convolutions with recurrent updates, so higher-order interactions and their temporal evolution are learned jointly. On the seven-class SEED-VII emotion recognition task it outperforms convolutional, recurrent, graph-based, and Transformer methods in both trial-wise and cross-subject evaluation, and incorporating eye-movement features improves it further — evidence for the framework’s potential in multimodal brain-state decoding.

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.