This repository provides the official implementation for the paper:

Complete-Tree Space Favors Data-Efficient Link Prediction

The codebase is organized as follows:

./ 
├── dataset/ 
├── logs/ 
├── special_req/ 
│ └── embedding_link_prediction_dw.py 
├── main.py 
├── utils.py 
├── split_datasets.py
├── synthetic_datasets.py
├── dataset_data_efficiency.sh 
├── dataset_practicality.sh
├── dataset_scalability.sh
├── exp_data_efficiency.sh
├── exp_practicality.sh 
└── exp_scalability.sh 

• dataset/: Where the datasets (like Cora) should be placed.
• logs/: Output directory for training and testing logs.
• special_req/: A modified file for CogDL’s embedding_link_prediction_dw.py.
• main.py: Entry point for main experiments.
• utils.py: Utility functions.
• split_datasets.py: Dataset splitter in our setting.
• bash scripts: Shell scripts to run different experiment scenarios.

1. Python Environment
   Make sure you have the following packages installed:

   • pytorch
   • scikit-learn
   • numpy
   • pandas
   • networkx
   • torch_geometric

2. Datasets
   We use CogDL for automatic dataset downloading. Please install CogDL in your environment

   pip install cogdl

   or refer to CogDL's GitHub repo for detailed installation instructions.

3. Replace the CogDL Link Prediction Wrapper In your local CogDL installation, find the file:

   cogdl/wrappers/data_wrapper/link_prediction/embedding_link_prediction_dw.py
   

   Replace its content with the file provided in our repository:

   ./special_req/embedding_link_prediction_dw.py
   

   This step ensures compatibility with our experimental setup.

We provide several pre-configured bash scripts to reproduce different experiment settings described in the paper. All results will be logged in ./logs/ for further analysis. As default, all the metrics including roc_auc, pr_auc, mrr, f1, hits20, hits50, hits100 would be report in every epoch for performance comparison. The loss of each epoch is reported for convergence judgement.

When splitting data, CogDL or torch_geometric may fail to download the dataset. Please manually download them in ./datasets.

To reproduce the data-efficiency experiments (varying $\mu$ from 0.1 to 0.9 on cora, citeseer, pubmed, icews18 and ogbl-collab):

bash ./dataset_data_efficiency.sh
bash ./exp_data_efficiency.sh

To assess the model’s performance on ogbl-collab and ogbl-ppa with $\mu = 0.02$:

bash ./data_practicality.sh
bash ./exp_practicality.sh

To evaluate scalability on synthetic graphs and the real graph (ogbl-collab):

bash ./dataset_scalability.sh
bash ./exp_scalability.sh