Scalable and Adaptive Graph Neural Networks with Self-Label-Enhanced training

This repository is an implementation of paper: Scalable and Adaptive Graph Neural Networks with Self-Label-Enhanced training

Requirements

pytorch dgl ogb pytorch-geometric

Statement

Feel free to use any parts from our works, but remember to cite our paper and repository. Introduce our methods if your work is mostly based on ours. We think these are basic academic ethics.

Datasets

Inductive datasets (Reddit, Flickr, Yelp) can be loaded via dgl or pyg. PPI should be separately downloaded and saved into dataset. More detailed information can be found in GraphSaint

OGB datasets are loaded via OGB library. More detailed information can be found in Open Graph Benchmark

Reproduce results

Run shell scripts on folder scripts, add device number to indicate GPU:

bash scripts/{dataset}/{method} {device_id}

Results

Test F1-micro with means and standard deviations under 10 runs. We use reported metrics for baselines.

Method	Reddit	Flickr	PPI	Yelp
GCN	93.3±0.0%	49.2±0.3%	51.5±0.6%	37.8±0.1%
FastGCN	92.4±0.1%	50.4±0.1%	51.3±3.2%	26.5±5.3%
Stochastic-GCN	96.4±0.1%	48.2±0.3%	96.3±1.0%	64.0±0.2%
AS-GCN	95.8±0.1%	50.4±0.2%	68.7±1.2%	—
GraphSAGE	95.3±0.1%	50.1±1.3%	63.7±0.6%	63.4±0.6%
ClusterGCN	95.4±0.1%	48.1±0.5%	87.5±0.4%	60.9±0.5%
GraphSaint	96.6±0.1%	51.1±0.1%	98.1±0.4%	65.3±0.3%
SGC	94.9±0.0%	50.2±0.1%	89.2±1.5%	35.8±0.6%
SIGN	96.8±0.0%	51.4±0.1%	97.0±0.3%	63.1±0.3%
SAGN	96.9±0.0%	51.8±0.2%	98.0±0.1%	65.4±0.1%
SAGN+SLE	97.1±0.0%	55.1±0.1%	98.1±0.1%	65.6±0.1%

Validation and test accuray with means and standard deviations under 10 runs. We use reported metrics for baselines.