/PLNLP

Pairwise learning neural link prediction for ogb link prediction

Primary LanguagePython

Pairwise Learning for Neural Link Prediction for OGB (PLNLP-OGB)

This repository provides evaluation codes of PLNLP for OGB link property prediction task. The idea of PLNLP is described in the following article:

Pairwise Learning for Neural Link Prediction (https://arxiv.org/pdf/2112.02936.pdf)

The performance of PLNLP on OGB link prediction tasks is listed as the following tables:

ogbl-ddi (Hits@20) ogbl-collab (Hits@50) ogbl-citation2 (MRR)
Validation 82.42 ± 2.53 100.00 ± 0.00 84.90 ± 0.31
Test 90.88 ± 3.13 70.59 ± 0.29 84.92 ± 0.29

Only with basic graph neural layers (GraphSAGE or GCN), PLNLP achieves top-1 performance on both ogbl-collab and ogbl-ddi, and top-2 on ogbl-citation2 in current OGB Link Property Prediction Leader Board until Dec 22, 2021 (https://ogb.stanford.edu/docs/leader_linkprop/), which demonstrates the effectiveness of the proposed framework. We believe that the performance will be further improved with link prediction specific neural architecure, such as proposed ones in our previous work [2][3]. We leave this part in the future work.

Environment

The code is implemented with PyTorch and PyTorch Geometric. Requirments:
 1. python=3.6
 2. pytorch=1.7.1
 3. ogb=1.3.2
 4. pyg=2.0.1

Reproduction of performance on OGBL

ogbl-ddi:

python main.py --data_name=ogbl-ddi --emb_hidden_channels=512 --gnn_hidden_channels=512 --mlp_hidden_channels=512 --num_neg=3 --dropout=0.3 

ogbl-collab:

Validation set is allowed to be used for training in this dataset. Meanwhile, following the trick of HOP-REC, we only use training edges after year 2010 with validation edges, and train the model on this subgraph. The performance of "PLNLP (val as input)" on the leader board can be reproduced with following command:

python main.py --data_name=ogbl-collab --predictor=DOT --use_valedges_as_input=True --year=2010 --train_on_subgraph=True --epochs=800 --eval_last_best=True --dropout=0.3

Furthermore, we sample high-order pairs with random walk and employ them as a kind of data augmentation. This augmentation method improves the performance significantly. To reproduce the performance of "PLNLP (random walk aug.)" on the leader board, you can use the following command:

python main.py --data_name=ogbl-collab  --predictor=DOT --use_valedges_as_input=True --year=2010 --train_on_subgraph=True --epochs=800 --eval_last_best=True --dropout=0.3 --gnn_num_layers=1 --grad_clip_norm=1 --use_lr_decay=True --random_walk_augment=True --walk_length=10 --loss_func=WeightedHingeAUC

ogbl-citation2:

python main.py --data_name=ogbl-citation2 --use_node_feat=True --encoder=GCN --emb_hidden_channels=50 --mlp_hidden_channels=200 --gnn_hidden_channels=200 --grad_clip_norm=1 --eval_steps=1 --num_neg=3 --eval_metric=mrr --epochs=100 --neg_sampler=local 

Reference

This work is based on our previous work as listed below:

[1] Zhitao Wang, Chengyao Chen, Wenjie Li. "Predictive Network Representation Learning for Link Prediction" (SIGIR'17) [Paper]

[2] Zhitao Wang, Yu Lei and Wenjie Li. "Neighborhood Interaction Attention Network for Link Prediction" (CIKM'19) [Paper]

[3] Zhitao Wang, Yu Lei and Wenjie Li. "Neighborhood Attention Networks with Adversarial Learning for Link Prediction " (TNNLS) [Paper]