sangyx / gtrick

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. It tests and provides powerful tricks to boost your models' performance.

Trick is all you need! (English Document | 中文介绍)

Library Highlights

• Easy-to-use: All it takes is to add a few lines of code to apply a powerful trick, with as little changes of existing code as possible.
• Verified Trick: All tricks implemented in gtrick are tested on our selected datasets. Only the tricks indeed improving model's performance can be collected by gtrick.
• Backend Free: We provide all tricks both in DGL and PyG. Whatever graph learning library you use, feel free to try it.

Installation

Note: This is a developmental release.

pip install gtrick

Quick Start

It is very easy to get start with gtrick. You can enhance your GNN model with only a few lines of code.

For more detailed examples, see Example in Trick.

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

We have tested all these tricks on selected datasets, you can find the benchmark in PyG Benchmark and DGL Benchmark.

Contributing

Please let me know if you encounter a bug or have any suggestions by filing an issue.

I welcome all contributions from bug fixes, new tricks and better examples.

There are also some trciks I'm planning to add, feel free to implement one of them for gtrick:

• SCR
• GSAT
• Edge Proposal Set

gtrick's Issues

Is the dimension of h and vx output by the 'update_node_emb' method in the 'VirtualNode' class consistent?

Hello, may I ask if the dimensions of h and vx output by the 'update_node_emb' method are the same? From the code perspective, it should be the same, but what I actually run is different. The number of lines in h is the same as the number of nodes, but the number of lines in vx is the same as "batch_size". Finally, it seems that due to the inconsistent dimensions of h and vx, the statement 'global_add_pool (x, batch)+vx' in the 'update_vn_emb' method reported an error.
The error reported is

RuntimeError: The size of tensor a (37) must match the size of tensor b (128) at non-singleton dimension 1

why should not use x += perturb

Hi, thanks for your work. In FLAG.py, you comment 'add perturb to x, note that do not use x += perturb' in line 65. And I want to know why should not use x += perturb.

why

边特征提示AttributeError: type object 'SparseTensor' has no attribute 'from_edge_index'

information about the environment

Hi,
Could you plz provide information about the environment in which the code was run (https://github.com/sangyx/gtrick/blob/main/benchmark/pyg/graph_pred.py), including the version of the dependencies used? This will help me to reproduce the results you obtained. Thank you very much.

When I try to use EGNN from model.py without adding VirtualNode

Why I get this kind of error: IndexError: index out of range in self
I'm trying to compare the result before and after adding a Virtual Node.

bug

https://github.com/sangyx/gtrick/blob/main/gtrick/dgl/virtual_node.py
Is line35 super(VirtualNode).__init__() wrong?

Question about VirtualNode

I want to add a virtual node to the graph, but I don't want to connect the whole graph, just some nodes in it. How should I do this?
For example: there is a node V={0,1,2,3,4} in the graph, I want to add a node s, connect {s,0},{s,2}, how to realise this is relatively simple.