gtrick 0.0.dev2

Bag of Tricks for Graph Neural Networks

gtrick: Bag of Tricks for Graph Neural Networks.

gtrick is an easy-to-use Python package collecting tricks for Graph Neural Networks. We test and provide powerful tricks to boost your models' performance.

Trick is all you need!(Chinese Introduction)

Trick

Trick	Example	Task	Reference
VirtualNode	DGL PyG	graph	OGB Graph Property Prediction Examples
FLAG	DGL PyG	node* graph	Robust Optimization as Data Augmentation for Large-scale Graphs
Fingerprint	DGL PyG	molecular graph*	Extended-Connectivity Fingerprints
Random Feature	DGL PyG	graph*	Random Features Strengthen Graph Neural Networks
Label Propagation	DGL PyG	node*	Learning from Labeled and Unlabeled Datawith Label Propagation
Correct & Smooth	DGL PyG	node*	Combining Label Propagation And Simple Models Out-performs Graph Neural Networks
Common Neighbors	DGL PyG	link*	Link Prediction with Structural Information
Resource Allocation	DGL PyG	link*	Link Prediction with Structural Information
Adamic Adar	DGL PyG	link*	Link Prediction with Structural Information
Anchor Distance	DGL PyG	link*	Link Prediction with Structural Information

Installation

Note: This is a developmental release.

pip install gtrick

Benchmark

The results listed below are implemented by PyG. You can find the results of DGL in DGL Benchmark.

Graph Property Prediction

Dataset	ogbg-molhiv
Trick	GCN	GIN
—	0.7690 ± 0.0053	0.7778 ± 0.0130
+Virtual Node	0.7581 ± 0.0135	0.7713 ± 0.0036
+FLAG	0.7627 ± 0.0124	0.7764 ± 0.0083
+Random Feature	0.7743 ± 0.0134	0.7692 ± 0.0065
Random Forest + Fingerprint	0.8218 ± 0.0022

Node Property Prediction

Dataset	ogbn-arxiv
Trick	GCN	SAGE
—	0.7167 ± 0.0022	0.7167 ± 0.0025
+FLAG	0.7187 ± 0.0020	0.7206 ± 0.0013
+Label Propagation	0.7212 ± 0.0006	0.7197 ± 0.0020
+Correct & Smooth	0.7220 ± 0.0037	0.7264 ± 0.0004

Link Property Prediction

Dataset	ogbn-collab
Trick	GCN	SAGE
—	0.4718 ± 0.0093	0.5140 ± 0.0040
+Common Neighbors	0.5332 ± 0.0019	0.5370 ± 0.0034
+Resource Allocation	0.5024 ± 0.0092	0.4787 ± 0.0060
+Adamic Adar	0.5283 ± 0.0048	0.5291 ± 0.0032
+AnchorDistance	0.4740 ± 0.0135	0.4290 ± 0.0107