Optimized Hyperparameters for Physics-Informed Graph Neural Networks

Hyperparameter values for an HP-GNN model were determined via systematic optimization including grid search and Bayesian ablation studies. The dataset, created by Masoud Amiri, details a specific scaling where the physics learning rate is set to 0.1 times the main rate for stable Kuramoto dynamics convergence. It was last updated in April 2026.

Use Cases

• Replicate the HP-GNN training setup by applying the documented physics learning rate scaling factor relative to the main learning rate.
• Analyze the effect of systematically optimized hyperparameters on model performance metrics like 'Chimera Detection Accuracy' or 'Gnn Achieved 84'.
• Benchmark new hyperparameter search methods against the results from grid search and Bayesian optimization documented for this model.
• Study the rationale behind hyperparameter choices for stabilizing the convergence of Kuramoto parameters in neural interaction models.

Strengths

• Hyperparameters were determined through systematic optimization methods including grid search and Bayesian optimization.
• Dataset is openly shared under a CC-BY-4.0 license, facilitating reuse and citation.
• Provides explicit rationale for key design choices, such as the physics learning rate scaling.

Limitations

• The dataset is very small at 9.5 KB, indicating limited scope and likely a single table of parameters without extensive supporting data.
• Sample data and specific column names are unavailable, hindering immediate understanding of the data structure.
• The model and parameters are tied to a specific application (seizure prediction with 22 pediatric patients), limiting generalizability.

Provenance

Source

figshare, authored by Masoud Amiri.

Collection Method

Values determined through systematic optimization (grid search, Bayesian optimization, ablation studies) on validation data.

Time Range

null

Freshness

Last updated in April 2026.

Geography

null