MLMP: Metapath-Enhanced Language Model Pretraining on Text-Attributed Heterogeneous Graphs

This repository contains the source code and datasets for MLMP: Metapath-enhanced Language Model Pretraining on Text-Attributed Heterogeneous Graphs.

Links

• Datasets
• Preprocess
• Pretraining
• Finetuning

Datasets

Download processed data. To reproduce the results in our paper, you need to first download the processed datasets. You need to also download bert-base-cased and put them into ./data.

Preprocess

You need to execute ./data/data_process.ipynb for OAG-Venue dataset and ./data/data_process_googreads.ipynb for GoodReads dataset.

Pretraining

Pretraining in ./pretrain.

sh run.sh

Finetuning

Node Classification

Run node classification in ./downstream/node-classification.

sh run.sh

Link Prediction

Run link prediction in ./downstream/link-predict.

sh run.sh