ProtonPrat/anlp-a3-gnn

ANLP A3 Part B: UPFD GossipCop graph classifiers

One checkpoint per configuration and seed, named <model>-<feature>-seed<k>.safetensors. Each holds the weights from the epoch with the best validation F1. Models are built with src.part2.models.build_model from the assignment repository, with the hyperparameters in config.yaml.

Test mean over seeds 0–4:

Experiment	Configuration	Accuracy	F1
architecture_comparison	root_mlp	0.7369	0.7313
architecture_comparison	gcn	0.9640	0.9642
architecture_comparison	graphsage	0.9668	0.9669
architecture_comparison	gat	0.9682	0.9683
architecture_comparison	bigcn	0.9545	0.9547
feature_ablation	profile	0.9573	0.9577
feature_ablation	spacy	0.9618	0.9618
feature_ablation	bert	0.9668	0.9669
feature_ablation	content	0.9725	0.9725