Role2Vec

A scalable parallel gensim implementation of Learning Role-based Graph Embeddings (IJCAI 2018).

Abstract

The second-order random walk sampling methods were taken from the reference implementation of Node2vec.

The model is now also available in the package Karate Club.

This repository provides an implementation of Role2Vec as described in the paper:

Learning Role-based Graph Embeddings. Nesreen K. Ahmed, Ryan Rossi, John Boaz Lee, Theodore L. Willke, Rong Zhou, Xiangnan Kong, Hoda Eldardiry. StarAI workshop - IJCAI, 2018. [Paper]

Requirements

The codebase is implemented in Python 3.5.2. package versions used for development are just below.

networkx          2.4
tqdm              4.28.1
numpy             1.15.4
pandas            0.23.4
texttable         1.5.0
scipy             1.1.0
argparse          1.1.0
gensim            3.6.0
scikit-learn      0.20.0

Datasets

Input and output options

  --graph-input      STR   Input graph path.   Default is `input/cora_edges.csv`.
  --output           STR   Embeddings path.    Default is `output/cora_role2vec.csv`.

Random walk options

  --window-size      INT    Skip-gram window size.        Default is 5.
  --walk-number      INT    Number of walks per node.     Default is 10.
  --walk-length      INT    Number of nodes in walk.      Default is 80.
  --sampling         STR    Sampling procedure.           Default is `first`.
  --P                FLOAT  Return parameter.             Default is 1.0.
  --Q                FLOAT  In-out parameter.             Default is 1.0.

Factorization options

  --dimensions      INT      Number of dimensions.      Default is 128
  --down-sampling   FLOAT    Down sampling frequency.   Default is 0.001.
  --alpha           FLOAT    Initial learning rate.     Default is 0.025.
  --min-alpha       FLOAT    Final learning rate.       Default is 0.025.
  --min-count       INT      Minimal feature count.     Default is 1
  --workers         INT      Number of cores.           Default is 4.
  --epochs          INT      Number of epochs.          Default is 10.

Feature creation options

  --features               STR     Feature extraction mechanism.         Default is `wl`.
  --labeling-iterations    INT     Number of WL labeling iterations.     Default is 2.
  --log-base               FLOAT   Log base for label creation.          Default is 1.5.
  --graphlet-size          INT     Maximal graphlet size.                Default is 4.
  --quantiles              INT     Number of quantiles for binning.      Default is 5.
  --motif-compression      STR     Motif compression procedure.          Default is `string`.
  --seed                   INT     Sklearn random seed.                  Default is 42.
  --factors                INT     Factors for motif compression.        Default is 8.
  --clusters               INT     Number of motif based labels.         Default is 50.
  --beta                   FLOAT   Motif compression regularizer.        Default is 0.01.

Examples

Using the degree centrality as a structural feature.

python src/main.py --features degree

Using the Weisfeiler-Lehman labeling as a structural feature.

python src/main.py --features wl

Using motif based structural features with factorization compression.

python src/main.py --features motif --motif-compression factorization

Using motif based structural features with factorization compression and a structural label number of 40.

python src/main.py --features motif --motif-compression factorization --clusters 40

Using a custom factorization dimension for the embedding.

python src/main.py --dimensions 32

Using second-order attributed ranom walks for sampling.

python src/main.py --sampling second --P 1 --Q 4

License

• GNU