tianjianl / arxiv-competition Goto Github PK

This repository contains a PaddlePaddle implementation of a Graph Neural Network for classification on a smaller ogbn-arxiv dataset.

If you want to see the results, you can find them in this notebook in Chinese, or this notebook in English.

in order to start training, simply type

sh train.sh

then all of the requirements should be installed automatically.
after the training loop, two .csv files will be generated. The two files submission.csv is the unsmoothed output and submission_cs.csv is the smoothed output of the model.

• graphmodel_1.pycontains the code for our graph neural network with residual connections.
• correctandsmooth.pycontains the code for the Correct & Smooth trick described in this paper
• unimpmodel.py contains the code for the Baidu UniMP model described in this paper

Combining Label Propagation and Simple Models Out-performs Graph Neural Networks https://arxiv.org/abs/2010.13993

Semi-Supervised Classification with Graph Convolutional Networks https://arxiv.org/abs/1609.02907

Masked Label Prediction: Unified Message Passing Model for Semi-Supervised Classification https://arxiv.org/abs/2009.03509