The question behind the work.
When node features are noisy but connected nodes tend to share labels, does message passing improve classification over an MLP that ignores adjacency?
Dataset & environment
Synthetic homophilic graph of 240 nodes, 3 classes, 16-dimensional noisy features and 2,148 undirected edges.
From the idea to the method.
- Generate a seeded synthetic homophilic graph with noisy node attributes and hidden class labels.
- Normalize the adjacency matrix using symmetric degree normalization and self-loops.
- Implement a two-layer GCN and a feature-only two-layer MLP baseline in PyTorch.
- Construct disjoint, class-balanced node splits for training, validation and testing.
- Fit models with Adam and compare held-out node accuracy on exactly this synthetic setup.
What was implemented.
- Implemented GCN aggregation, normalized adjacency and a feature-only MLP baseline.
- Implemented seeded train/validation/test splits and validation-based checkpoint selection.
How the work is evaluated.
Compare accuracy on identical held-out nodes. Validate adjacency symmetry and split disjointness; note that this small synthetic benchmark favors structural information.
What is documented.
On one deliberately homophilic synthetic graph (seed 31): GCN test accuracy 100%; MLP test accuracy 54.49%. These results illustrate message passing in this specific synthetic setup only; they are not standard benchmark or real-world performance claims.
Synthetic, seed-controlled experiments. Not a standard benchmark and not evidence of performance on real datasets.
A foundation for the next experiment.
- Graph structure may carry predictive signal that is absent from raw node features.
- A synthetic controlled comparison is useful for studying inductive biases but is not a substitute for public benchmark evaluation.