GNNBenchmarkValidator<T>

Models & Types · AiDotNet.Validation

Validates GNN implementations against expected benchmarks and behaviors.

For Beginners

Why benchmark validation matters:

Purpose of Validation:

  • Ensure implementations are correct
  • Compare against published results
  • Detect regressions when code changes
  • Verify performance claims

Common Validation Checks:

1. Sanity Checks:

  • Model can overfit small dataset (proves it can learn)
  • Predictions change after training (proves parameters update)
  • Gradients flow correctly (no vanishing/exploding)

2. Architecture Tests:

  • Layer output shapes correct
  • Adjacency matrix handling works
  • Pooling produces fixed-size outputs

3. Benchmark Comparisons:

  • Cora node classification: Should reach ~80% accuracy
  • ZINC generation: Should produce valid molecules
  • Link prediction AUC: Should beat random (>0.5)

4. Invariance Tests:

  • Node permutation invariance (reordering shouldn’t change result)
  • Edge order independence
  • Batch independence

Example Usage:

var validator = new GNNBenchmarkValidator<double>();

// Test node classification
var nodeResults = validator.ValidateNodeClassification();
// Result is available in the returned value

// Test graph classification
var graphResults = validator.ValidateGraphClassification();
// Result is available in the returned value

How It Works

This class provides validation methods to ensure GNN implementations meet academic standards and reproduce expected results on standard benchmarks.

Constructors

Constructor Summary
GNNBenchmarkValidator()

Methods

Method Summary
Dictionary<string, double> ValidateGraphGeneration() Validates graph generation produces valid molecules.
GraphClassificationResults<T> ValidateGraphClassification(MolecularDataset<T> datasetType) Validates graph classification on molecular dataset.
LinkPredictionResults<T> ValidateLinkPrediction() Validates link prediction on citation network.
NodeClassificationResults<T> ValidateNodeClassification(CitationDataset<T> datasetType) Validates node classification on citation network.
bool ValidatePermutationInvariance() Validates permutation invariance (node order shouldn’t matter).