Lab 1 — Your first GNN: GCN on Cora

Goal. Train a 2-layer GCN on the Cora citation network and compare against a logistic-regression baseline on the same features.

What you ship. Notebook with training curves, final test accuracy for both models, and a 200-word memo on what the GCN gains over the bag-of-features baseline.

Setup

Install the dependencies (one-time).

# !pip install torch torch_geometric scikit-learn matplotlib

import torch
import torch.nn.functional as F
from torch_geometric.datasets import Planetoid
from torch_geometric.nn import GCNConv
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score
import numpy as np

torch.manual_seed(42)
np.random.seed(42)

The Cora citation network

dataset = Planetoid(root='/tmp/Cora', name='Cora')
data = dataset[0]
print('Nodes:', data.num_nodes, 'Edges:', data.num_edges, 'Features:', data.num_features, 'Classes:', dataset.num_classes)

Exercise 1 — Logistic-regression baseline on node features

X = data.x.numpy()
y = data.y.numpy()
train_mask = data.train_mask.numpy()
test_mask = data.test_mask.numpy()

clf = LogisticRegression(max_iter=2000)
clf.fit(X[train_mask], y[train_mask])
preds = clf.predict(X[test_mask])
print('Logistic baseline accuracy:', accuracy_score(y[test_mask], preds))

Exercise 2 — Implement a 2-layer GCN

# YOUR TURN
# Define a GCN with two GCNConv layers, a ReLU between them, and dropout.
# class GCN(torch.nn.Module): ...

Exercise 3 — Train and report test accuracy

# YOUR TURN
# Train for 200 epochs with Adam(lr=0.01). Track val accuracy.
# Report final test accuracy. Compare to the logistic baseline.

Exercise 4 — Inspect failure cases

# YOUR TURN
# Find 5 test-set nodes the GCN gets wrong. Look at their neighborhoods.
# Hypothesize why and write a short note.

Done?

Submit per the cohort schedule. Peer review pairing announced the following Monday.