meotecarlo / rgat

A TensorFlow implementation of Relational Graph Attention Networks, paper: https://arxiv.org/abs/1904.05811

Relational Graph Attention Networks

A TensorFlow implementation of Relational Graph Attention Networks for semi-supervised node classification and graph classification tasks introduced in our paper Relational Graph Attention Networks. It is compatible with static and eager execution modes.

Contact [email protected] for comments and questions.

Installation

To install rgat, run:

$ pip install git+git://github.com/Babylonpartners/rgat.git

To allow both the CPU and GPU versions of TensorFlow, we have not listed Tensorflow as a requirement. You will need to install a version separately. The TensorFlow version must be >= 1.10.

Examples

To make it easy to jump right in, we have provided a few examples:

• RDF: Semi-supervised node classification on AIFB and MUTAG.
• Batch comparison: Batched graph forward pass examples using both static and eager mode.

Usage

The graphs handles are composed of:

• An inputs dense tensor* corresponding to node features, and
• A support sparse tensor corresponding to node connectivity.

Below is a standard pattern for using RGAT

from rgat.layers import RGAT

inputs = get_inputs()                                # Dense tensor with shape (?, Features)

support = get_support()                              # Sparse tensor with dense shape (?, ?)
support = tf.sparse_reorder(support)                 # May be neccessary, depending on construction

rgat = RGAT(units=FLAGS.units, relations=RELATIONS)  # RELATIONS is an integer indicating the number 
                                                     # of relation types in the graph

outputs = rgat(inputs=inputs, support=support)       # Dense tensor with shape (?, FLAGS.units)

Here, get_inputs and get_support are user-provided functions. For guidance on their definition, see the usage guide.

*Note, it is possible to pass a SparseTensor as the inputs to the layer call. This will result in the identity matrix for the input (but skipping the trivial matrix multiply).

Cite

Please cite our paper if you use this code in your own work:

@article{busbridge2019rgat,
title = {Relational Graph Attention Networks},
author = {Busbridge, Dan and Sherburn, Dane and Cavallo, Pietro and Hammerla, Nils Y},
year = {2019},
eprint = {arXiv:1904.05811},
url = {http://arxiv.org/abs/1904.05811}
}