HetLGN

A PyTorch implementation for the CIC 2020 paper below:
Discovering Localized Information for Heterogeneous Graph Node Representation Learning.
Lin Meng^*, Ning Yan, Masood Mortazavi, Jiawei Zhang.
[paper]

^*This work is based on the summer internship with Futurewei, Inc.

Dataset

We test it on three datasets, the preprocessed datasets are in ./data/:

• ACM: graph_acm_v3.pk, labels_acm_v3.pk
• DBLP: graph_DBLP_four_area.pk, labels_DBLP_four_area.pk
• IMDB: graph_imdb.pk, labels_imdb.pk

All raw datasets are publicly accessible in the page.

Our code reuses part of codes in HGT

Requirements

Codes are written under python 3.7

• pytorch_geometric 1.6.1+
• pytorch 1.5.0+
• scipy
• numpy
• pandas
• skicit-learn
• dill

If using docker, please pull pytorch/pytorch:latest image and install the required packages.

Run

The preprocessing codes are files with name starting with preprocess.

To run the code, simply use '''python preprocess_xxx.py'''

'''bash python train_node_classification.py --sample_width 6 --sample_depth 2 --dataset _acm_v3 --target_type paper --n_hid 8 > acm.out

python train_node_classification.py --sample_width 5 --sample_depth 2 --dataset _DBLP_four_area --target_type author --n_hid 4 > dblp.out

python train_node_classification.py --sample_width 4 --sample_depth 2 --dataset _imdb --target_type movie --n_hid 8 > imdb.out 

'''

Docker environment set up

If you prefer to use docker, please cd to the dockerfile folder /home/meng/code/isonode .

• RUN dockerfile to build the required image '''bash docker build --tag pygraph:1.0 '''

• RUN a docker container '''bash docker run --gpus all -d --ipc=host -it pygraph:1.0 /bin/bash '''

• Then got the CONTAINER_ID by '''docker ps''' and enter the container via bash '''bash docker exec -it CONTAINER_ID /bin/bash '''

Plot

The plot file is node_clustering.py