Skip to content

Latest commit

 

History

History
104 lines (89 loc) · 4.29 KB

README.md

File metadata and controls

104 lines (89 loc) · 4.29 KB
Raw

Graph Embeddings

A package for creating embedding vectors for nodes in a relational graph. Input a CSV file (or a directory of CSVs) of connections in a graph (as well as optional features of the nodes of the graph), and receive a list of (node_name, node_embedding) tuples for each node in the graph.

Installation

This package was created with Python 2.7 and works with Keras 2.2.4 using Tensorflow 1.12.0 as the backend. It's recommended to use Anaconda to set up a clean environment in order to ensure no conflicts between various packages. You can download the anaconda package manager here: https://www.anaconda.com/download/

  1. Create the anaconda enviorment
conda create --name gembed python=2.7
conda activate gembed
  1. Download the repo
git clone https://github.com/bhaney/graph_embeddings
cd graph_embeddings
  1. Install all of the dependencies
pip install git+https://github.com/bhaney/relational-gcn.git
python setup.py install
  1. Run the program using the AIFB test data
cd gembed
python get_graph_embeddings.py -a rgcn -p data/aifb/relations -t data/aifb/person_affiliations.csv -e 50 -d 16 -n rgcn
python get_graph_embeddings.py -a ae -p data/aifb/relations -e 1 -d 16 -n autoencode
python get_graph_embeddings.py -a distmult -p data/aifb/relations -e 10 -d 16 -n distmult

The autoencoder needs about 50 epochs to produce a good embedding, but that can take more than an hour.

  1. To plot the AIFB embeddings, there is an iPython notebook you can use.
conda install jupyter matplotlib
conda install -c conda-forge umap-learn
jupyter notebook

Then select plot_aifb_graph_embeddings.ipynb

Usage

The program get_graph_embeddings.py takes the following arguments:

  -p PATH,      --path PATH,          Provide path to the directory of .CSV files.
  -i INPUT,     --input INPUT,        If you only want to run over one file, give path of the single CSV file..

  -n NAME,      --name NAME,          Name of the output file of the embeddings in the `results/` directory.
  -a ALGO,      --algo ALGO,          Which algorithm to use.
  -d DIM,       --dim DIM,            Desired embedding dimension.
  OPTIONAL:
  -e EPOCHS,    --epochs EPOCHS,      Number of epochs to train. Default is 1.
  -t TARGET,    --target TARGET,      CSV file with targets for training.
  -f FEATURES,  --features FEATURES,  JSON file with features for training.
    ,           --eigen EIGEN,        Number of eigenvectors to use in spectral analysis.
  -u,           --undirected,         Flag for undirected graphs. Default is directed.

The CSV file has to represent the graph by triplet connection of subject_node, relation_edge, predicate_node. Each row should contain one connection.

Example:

dog,IS,mammal
dog,MAKES_SOUND,woof
cat,IS,mammal
cat,MAKES_SOUND,meow
mouse,IS,mammal
mouse,MAKES_SOUND,squeak
ball,IS,toy
toy,MAKES_SOUND,squeak
mammal,HAS,fur
rufus,INSTANCE_OF,dog
spot,INSTANCE_OF,dog
algernon,INSTANCE_OF,mouse
rufus,LOCATED_IN,chicago
spot,LOCATED_IN,chicago
chicago,IS,city

The output of get_graph_embeddings will be a list of (node_name, node_embedding) for each node in the graph. It you run it from the command line, the embeddings will be saved as a .pkl file in the results/ directory.

Multigraph Summary

To see a summary of facts about the graph, you can use the summary() function from the Multigraph class. As an example:

python
>>> from gembed.multigraph import get_graph
>>> g = get_graph('gembed/data/example_graph.csv')
Processed 15 lines.
>>> g.summary()
n nodes: 15
n relation types: 5
n connections: 15
5 most connected nodes: [(u'dog', 2), (u'cat', 2), (u'mouse', 2), (u'rufus', 2), (u'spot', 2)]
5 least connected nodes (non-terminal): [(u'mammal', 1), (u'ball', 1), (u'toy', 1), (u'algernon', 1), (u'chicago', 1)]
n terminal nodes: 5
5 most frequent relations: [(u'IS', 5), (u'MAKES_SOUND', 4), (u'INSTANCE_OF', 3), (u'LOCATED_IN', 2), (u'HAS', 1)]
5 least frequent relations: [(u'IS', 5), (u'MAKES_SOUND', 4), (u'INSTANCE_OF', 3), (u'LOCATED_IN', 2), (u'HAS', 1)]
equivalent nodes: [[u'rufus', u'spot']]

Acknowledgements

The relation-GCN algorithm is taken from an implementation of Modeling Relational Data with Graph Convolutional Networks (2017) by Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling.