| Abstract: |
| Temporal graph classification models time evolving networks in neuroscience, cybersecurity, and infrastructure. Existing temporal GNNs often miss global structural evolution and can be sensitive to noise and node permutations. In this talk, we present \textbf{T3former}, a topological machine learning framework that augments temporal graph models with stable global descriptors. T3former extracts spectral summaries and persistent homology features from each snapshot to capture multi scale dynamics such as connectivity changes and the emergence or disappearance of cycles. These signals are fused with temporal neural architectures to produce robust sequence level representations. Across real world datasets including brain, social, and traffic networks, T3former improves classification performance over strong TGNN baselines, especially when labels depend on structural dynamics. |
|