DirLinkBench

This repository contains the implementation of our preprint paper "Rethinking Link Prediction for Directed Graphs".

Environment Setup

Ensure your environment meets the following dependencies:

• Python 3.10.13
• PyTorch 2.0.1
• PyTorch-CUDA 11.7
• torch-geometric 2.4.0
• NumPy 1.26.3
• SciPy 1.11.4
• scikit-learn 1.3.0
• ogb 1.3.6
• pandas 2.1.4
• DGL 0.9.0 (only required for DUPLEX)
• TensorBoard 2.10.1 (only required for ELTRA)
• TensorFlow 2.10.0 (only required for ELTRA)

Datasets

This repository includes two small datasets: Cora-ML and CiteSeer. Additional datasets can be downloaded automatically when needed.

Running Experiments

Baselines

To run baselines such as MLP, GCN, GAT, APPNP, GPRGNN, DGCN, DiGCN, DiGCNIB, DirGNN, MagNet, DUPLEX, DHYPR, and DiGAE, use the following command:

python ./scripts/name.py --dataset data_name

• Replace name.py with the script name (e.g., gcn.py, dirgnn.py).
• Replace data_name with the dataset name (e.g., cora_ml, citeseer).
• This command automatically searches all parameters for baselines and stores the results in the ./results folder.

STRAP, ELTRA, and ODIN

These methods require generating embeddings before running the link prediction task.

ODIN

python -m odin.main_embedding --dataset data_name
python -m odin.main_lp --dataset data_name --predictor Logistic --operator cat

ELTRA

python -m eltra.main_embedding --dataset data_name
python -m eltra.main_lp --dataset data_name --predictor Logistic --operator cat

STRAP

STRAP requires converting the graph data into .txt format before generating embeddings using the official STRAP repository:

python -m strap.data_transform --dataset data_name
python -m strap.main_lp --dataset data_name --predictor Logistic --operator cat

• --predictor options: Logistic, InnerProduct
• --operator options: cat, hadamard, cat2

Our SDGAE

To search for optimal parameters for SDGAE, run:

python ./scripts/sdgae.py --dataset data_name

To reproduce directly the results in Table 4, run:

bash ./scripts/sdgae.sh