Node Classification Eval

node-classification-eval

Interpretable and Adaptive Node Classification on Heterophilic Graphs via Combinatorial Scoring and Hybrid Learning — Vahidi (2025) (arXiv:2512.22221, 2025)

What this evaluates

This evaluation probes a model's ability to perform node classification on graphs across a spectrum of homophily regimes, from strongly heterophilic to strongly homophilic. It specifically tests the framework's adaptive capability to switch between a combinatorial predictor and a neural refinement stage based on validation performance.

Datasets

• Texas — total 183; splits: train (-1), val (-1), test (-1); repo https://github.com/AhmedBegggaUA/Diffusion-Jump-GNNs/tree/main/splits
• Cornell — total 183; splits: train (-1), val (-1), test (-1); repo https://github.com/AhmedBegggaUA/Diffusion-Jump-GNNs/tree/main/splits
• Actor — total 7600; splits: train (-1), val (-1), test (-1); repo https://github.com/AhmedBegggaUA/Diffusion-Jump-GNNs/tree/main/splits
• CiteSeer — total 3327; splits: train (-1), val (-1), test (-1); repo https://github.com/AhmedBegggaUA/Diffusion-Jump-GNNs/tree/main/splits
• Cora — total 2708; splits: train (-1), val (-1), test (-1); repo https://github.com/AhmedBegggaUA/Diffusion-Jump-GNNs/tree/main/splits
• Pubmed — total 19717; splits: train (-1), val (-1), test (-1); repo https://github.com/AhmedBegggaUA/Diffusion-Jump-GNNs/tree/main/splits

Metrics

• classification accuracy (primary) — range: percent
  • Fraction of correctly predicted node labels out of the total number of nodes in the test set. Results are averaged across 10 predefined data splits (48% train / 32% val / 20% test), with standard deviation reported.

Input / output format

Input: Graph adjacency structure, node feature vectors, and node labels (for training/validation). The model processes the graph structure and features to predict class labels for each node.

Output: A predicted class label for each node in the graph.

Scoring recipe

def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return (correct / len(gold_labels)) * 100

# Average over 10 fixed splits
split_accuracies = [compute_accuracy(preds[i], gold[i]) for i in range(10)]
mean_accuracy = sum(split_accuracies) / 10
std_accuracy = (sum((x - mean_accuracy)**2 for x in split_accuracies) / 10) ** 0.5

Common pitfalls

• Using random or standard 80/10/10 splits instead of the 10 fixed predefined splits (48/32/20) explicitly required for comparability.
• Tuning hyperparameters on validation or test nodes, which violates the protocol's requirement for training-only cross-validation/random search.
• Applying the neural refinement stage unconditionally; the protocol mandates a validation-gated selection where refinement is only used if it improves validation accuracy by a fixed margin.

Evidence (verbatim from paper)

All results are reported as classification accuracy averaged over the 10 splits, together with the corresponding standard deviation. For each split, model selection and all validation-based decisions are performed exclusively on the validation set, and final accuracy is reported only on the held-out test set.

Citation

@misc{vahidi2025interpretable,
  title={Interpretable and Adaptive Node Classification on Heterophilic Graphs via Combinatorial Scoring and Hybrid Learning},
  author={Vahidi (2025)},
  year={2025},
  note={arXiv:2512.22221}
}

• arXiv: 2512.22221