Abstract:
In this thesis, we investigated the general framework of Graph Neural Network for
node and graph classification tasks. We studied the phenomenon of over smoothing
and conducted experiments to validate them. We propose a novel spectral-based
Legendre Filter based on the Legendre polynomial to learn node features on graphstructured
data. We also described various aggregation schemes that can be employed
with the Legendre Filter to further improve information aggregation. Furthermore,
we proposed a novel algorithm that changes graph topology based on some heuristics
to improve overall classification accuracy. For Semi-Supervised learning, we demonstrated
that our proposed method performs better than GCN and Chebyshev Filter
on Citation Network datasets. Our proposed model outperformed GAT on Citeseer
and PubMed. For the full-supervised learning task, we showed that our method outperforms
all three baselines; GCN, GAT, and Chebyshev Filter on Citation Network
and WebKB datasets. We further showed that our method outperforms deep GNN
models like GCNII, JKNet on the WebKB dataset.
