PT_propagation_then_training

This is the official pytorch implementation of Propagation then Training (PT), a graph algorithm. PT is proposed in the paper:

On the Equivalence of Decoupled Graph Convolution Network and Label Propagation

by Hande Dong, Jiawei Chen, Fuli Feng, Xiangnan He, Shuxian Bi, Zhaolin Ding, Peng Cui

Published at WWW 2021.

Introduction

In this work, we propose a two-step label propagation algorithm: (1) propagating the label of training set, (2) training a base predictor supervised by the soft label obtained in the previous step. Through gradient analysis, we proved that it is equivalent to decoupled GCN.

Environment Requirement

The code runs well under python 3.8.5. The required packages are as follows:

• pytorch == 1.6.0
• numpy == 1.19.1
• scipy == 1.5.2

Datasets

Following APPNP, we use four datasets: CITESEER, CORA_ML, PUBMED and MS_ ACADEMIC.

CITESEER, CORA_ML, and PUBMED are citation networks, where each node represents a paper and an edge indicates a citation relationship. MS_ACADEMIC is a co-authorship network, where nodes and edges represent authors and co-author relationship, respectively.

Run the Code

To run Propagation then Training Adaptively (PTA), execute the following command:

python train_PT.py --dataset cora_ml --K 10 --alpha 0.1 --dropout 0.0 --epochs 2000 --hidden 64 --lr 0.1 --weight_decay 0.005 --loss_decay 0.05 --fast_mode False --mode 2 --epsilon 100 --str_noise_rate 2.0 --lbl_noise_num 0 --patience 100

To run Propagation then Training Statically (PTS, a variant of PT), execute the following command:

python train_PT.py --dataset cora_ml --K 10 --alpha 0.1 --dropout 0.0 --epochs 2000 --hidden 64 --lr 0.01 --weight_decay 0.005 --loss_decay 0.05 --fast_mode False --mode 0 --epsilon 100 --str_noise_rate 2.0 --lbl_noise_num 0 --patience 100

To run Propagation then Training Dynamically (PTD, a variant of PT), execute the following command:

python train_PT.py --dataset cora_ml --K 10 --alpha 0.1 --dropout 0.0 --epochs 2000 --hidden 64 --lr 0.1 --weight_decay 0.005 --loss_decay 10 --fast_mode False --mode 1 --epsilon 100 --str_noise_rate 2.0 --lbl_noise_num 0 --patience 100

To run APPNP (our main baseline), execute the following command:

python train_APPNP.py --dataset cora_ml --K 10 --alpha 0.1 --dropout 0.5 --epochs 2000 --hidden 64 --lr 0.01 --weight_decay 0.005 --str_noise_rate 2.0 --lbl_noise_num 0 --patience 100

• Above commands are for Cora_ML dataset. For Citeseer or Pubmed, replace '--dataset cora_ml' with '--dataset citeseer' or '--dataset pubmed'. For MS_Academic, replace '--dataset cora_ml' with '--dataset ms_academic' and replace '--alpha 0.1' with '--alpha 0.2'.
• PTS, PTD, and PTA are three variants of PT. The argument '--mode' is for the three variants by setting 0/1/2. PTA performs best among them.
• The argument '--str_noise_rate' is for specifying structure noise rate in graph. To keep the original graph, set '--str_noise_rate' to 2.0.
• The argument '--lbl_noise_num' is for specifying label noise numbers of training set. To keep the original labels, set '--lbl_noise_num' to 0.

Contact

Please contact [email protected] if you have any questions about the code and paper.

Reference

This code refers to:

https://github.com/klicperajo/ppnp

https://github.com/tkipf/gcn