SLPE

This repository is the official PyTorch implementation of "Understanding and Improving Laplacian Positional Encodings For Temporal GNNs", featuring temporal graph neural networks with trajectory-based supra-Laplacian positional encoding.

📄 Paper: Understanding and Improving Laplacian Positional Encodings For Temporal GNNs (ECML-PKDD 2025)

Overview

SLPE introduces novel positional encoding techniques for temporal graphs by leveraging trajectory information in supra-adjacency matrices. The framework extends traditional graph neural networks to handle dynamic graphs more effectively by incorporating temporal structural information through eigenvector trajectories.

Key Features

• Temporal Positional Encoding: Novel approach using supra-Laplacian eigendecomposition with trajectory information
• Multiple Graph Representations: Support for Regular, Concatenated, and Supra graph representations
• Flexible Model Architecture: Integration with HTGN, SLATE, and other temporal GNN models
• Comprehensive Dataset Support: Compatible with TGBL, DyGLib, and HTGN datasets
• GPU Acceleration: CUDA support with CuPy for efficient eigendecomposition

Installation

Environment Setup

Create a conda environment using the provided configuration:

conda env create -f environment.yml
conda activate env

Usage

Quick Start

Run the basic training script:

python src/train.py

Configuration

The framework supports extensive configuration through command-line arguments. Key parameters include:

Dataset Options

• --dataset: Choose from TGBL datasets (tgbl-wiki, tgbl-review, tgbl-coin, etc.), DyGLib datasets (CanParl, USLegis, Flights, etc.), or HTGN datasets (enron10, dblp, as733, etc.)

Model Architecture

• --model: Model type (default: HTGN)
• --nhid: Hidden embedding dimension (default: 16)
• --nfeat: Input feature dimension (default: 128)

Temporal Positional Encoding

• --use_pes: Enable positional encoding features
• --pes_kind: Type of PE (Laplacian, Random, Sinusoidal)
• --graph_representation: Graph representation method (Regular, Concatenated, Supra)
• --length_of_pe: Dimensionality of positional encoding (default: 8)
• --pes_window_size: Temporal context window size (default: 3)
• --use_traj: Enable trajectory-based eigenvector computation
• --circular: Use circular temporal connections

Training Parameters

• --max_epoch: Maximum training epochs (default: 500)
• --lr: Learning rate (default: 0.01)
• --patience: Early stopping patience (default: 50)
• --repeat: Number of experimental runs (default: 1)

Example Commands

Train HTGN with supra-Laplacian positional encoding:

python src/train.py --model=HTGN --dataset=enron10 --use_pes=True --pes_kind=Laplacian --graph_representation=Supra --use_traj=True

Train SLATE model on TGBL dataset:

python src/train.py --model=SLATE --dataset=tgbl-wiki --use_pes=True --length_of_pe=16 --pes_window_size=5

Supported Datasets

TGBL Datasets

• tgbl-wiki, tgbl-review, tgbl-coin, tgbl-flight, tgbl-comment

DyGLib Datasets

• CanParl, USLegis, Flights, UNtrade, UNvote, Contacts

HTGN Datasets

• enron10, dblp, as733, fbw, HepPh30

Architecture

Core Components

1. TemporalPE: Handles temporal positional encoding computation
2. EigenSolver: Manages eigendecomposition with multiple solver options
3. Runner: Orchestrates training and evaluation procedures
4. Model Implementations: HTGN, SLATE, and other temporal GNN variants

Graph Representation Methods

• Regular: Standard snapshot-based processing
• Concatenated: Feature concatenation across time windows
• Supra: Supra-adjacency matrix construction with inter-layer connections

Experiment Tracking

The framework integrates with Weights & Biases for comprehensive experiment tracking. Results include:

• AUC and AP scores across multiple runs
• Training curves and performance metrics
• Hyperparameter configurations
• Computational profiling information

Contributing

This project builds upon several existing frameworks:

• HTGN - Hyperbolic Temporal Graph Networks

Citation

If you use this code in your research, please cite:

@article{galron2025understanding,
  title={Understanding and Improving Laplacian Positional Encodings For Temporal GNNs},
  author={Galron, Yaniv and Frasca, Fabrizio and Maron, Haggai and Treister, Eran and Eliasof, Moshe},
  journal={arXiv preprint arXiv:2506.01596},
  year={2025}
}

License

Please refer to the individual component licenses and the original HTGN repository for licensing information.