jl-metric 0.1.1

A Johnson-Lindenstrauss-based metric for dynamic graph generative model evaluation

Quality Measures for Dynamic Graph Generative Models

This repository contains the source code for our manuscript Quality Measures for Dynamic Graph Generative Models. Our JL-Metric provides a principled approach to evaluating generative models for continuous-time dynamic graphs (CTDGs). This repository contains both the metric implementation for practical use and the experimental code from our paper.

Quick Start

To evaluate the quality of your generated dynamic graphs, follow the instructions below:

Installation

pip install jl-metric

Basic Usage

Our evaluator accepts graphs in the form of PyTorch Geometric TemporalData objects. These should contain temporal interactions with source nodes (src), destination nodes (dst), timestamps (t), and optionally messages/features (msg).

from jl_metric import JLEvaluator

# Initialize the evaluator
evaluator = JLEvaluator()

# Prepare your graphs as PyG TemporalData objects
# reference_graph = your ground truth or reference graph
# generated_graph = your model's generated graph

# Create input dictionary
input_dict = {
    'reference': reference_graph,
    'generated': generated_graph
}

# Evaluate and get results
result_dict = evaluator.eval(input_dict)
print(f"JL-Metric score: {result_dict['JL-Metric']}")

Advanced Configuration

The evaluator accepts several optional parameters:

evaluator = JLEvaluator(
    node_dim=100,     # Dimension for node embeddings
    graph_dim=100,    # Dimension for graph embeddings
    seed=42           # Random seed for reproducibility
)

Paper Experiments

The experiments from our paper can be reproduced using the code in the experiments/ directory:

# Install additional requirements for experiments
cd experiments
pip install -r requirements.txt

# Follow experiment-specific README
cat README.md

For details on our experimental methodology and to reproduce the results from our paper, please refer to the experiments README.

Citing Our Work

If you use JL-Metric in your research, please cite our paper:

@inproceedings{jl-metric,
    title={Quality Measures for Dynamic Graph Generative Models},
    author={Ryien Hosseini and Filippo Simini and Venkatram Vishwanath and Rebecca Willett and Henry Hoffmann},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
    url={https://openreview.net/forum?id=8bjspmAMBk}
}