🦅 EAGLE

Official code for the paper:

When Speed meets Accuracy: an Efficient and Effective Graph Model for Temporal Link Prediction

📦 Requirements

Required packages and versions:

• matplotlib==3.8.4
• networkx==3.3
• numba==0.60.0
• numpy==1.26.4
• ogb==1.3.6
• pandas==1.5.3
• pyg-lib==0.4.0+pt20cu118
• scikit-learn==1.5.2
• scipy==1.13.1
• torch==2.0.1+cu118
• torch_geometric==2.5.3

We can install the required dependencies with:

pip install -r requirements.txt

---

📊 Datasets

🔗 Link Prediction

We conducted experiments on the Contacts, LastFM, Wikipedia, Reddit, AskUbuntu, SuperUser, and Wiki-Talk datasets. The raw dataset files can be obtained from the following link:

Google Drive Link

Please place them under ./data/ before running codes.

Note: For our experimental setup, we generated 99 negative samples for each sample in the test set. For details on data processing, please refer to our code.

🎯 Node Classification

We conducted experiments on the Trade, Genre, Reddit, and Token datasets, using the raw dataset files and splits provided by TGB Datasets.

To install the TGB dependency, run:

pip install py-tgb

We recommend using the TGB's interface, which includes downloading and standardized splitting.

For further details, please consult the TGB documentation.

---

▶️ Run & Eval EAGLE

🔗 Link Prediction

🔗 EAGLE-Structure

To run EAGLE-Structure, use the following command:

cd link_prediction

python train_structure.py --dataset_name wikipedia --topk 100 --alpha 0.9 --beta 0.8 --gpu 0

The optimal parameters are as follows:

Dataset	topk	$\alpha$ in Equation (8)	$\beta$ in Equation (9)
Contacts	100	0.3	0.3
LastFM	100	0.9	0.2
Wikipedia	100	0.9	0.8
Reddit	100	0.9	0.9
AskUbuntu	50	0.3	0.5
SuperUser	50	0.2	0.5
WikiTalk	100	0.5	0.5

🔗 EAGLE-Time

To run EAGLE-Time, use the following command:

cd link_prediction

python train_time.py --dataset_name wikipedia --topk 15 --lr 0.001 --weight_decay 5e-5 --gpu 0

The optimal parameters are as follows:

Dataset	top-kr in Equation (6)	learning_rate	weight_decay
Contacts	30	0.0001	5e-05
lastfm	20	0.0001	5e-05
wikipedia	15	0.001	5e-05
reddit	50	0.001	0.0
askubuntu	30	0.0001	0.0
superuser	30	0.0001	0.0
wikitalk	30	0.001	0.0

🔗 EAGLE-Hybrid

To run EAGLE-Hybrid, use the following command:

cd link_prediction

python train_hybrid.py --dataset_name wikipedia --gpu 0

Note: EAGLE-Hybrid is a weighted combination of EAGLE-Structure and EAGLE-Time. You need to first run structure module and time module with their optimal parameters listed above before training EAGLE-Hybrid.

---

🎯 Node Classification

To run EAGLE, use the following command:

cd node_classification

python train.py --dataset_name tgbn-trade --k 50 --tppr_alpha 0.6 --tppr_beta 0.9 --gamma 0.9 --window 4 --gpu 0

The optimal parameters are as follows:

Dataset	topk	$\alpha$ in Equation (8)	$\beta$ in Equation (9)	self_weight $\gamma$	window_size
tgbn-trade	50	0.6	0.9	0.9	4
tgbn-genre	20	0.1	0.1	0.1	7
tgbn-reddit	20	0.1	0.1	0.1	6
tgbn-token	20	0.1	0.1	0.1	5

---

❤️ Thank you for your interest in our work! ❤️