An Empirical Evaluation of Network Representation Learning Methods

This repository contains the instructions and materials necessary for reproducing the experiments presented in the paper: An Empirical Evaluation of Network Representation Learning Methods

The repository is maintained by Alexandru Mara (alexandru.mara@ugent.be).

Reproducing Experiments

In order to reproduce the experiments presented in the paper the following steps are necessary:

1. Download and install the EvalNE library v0.3.3 as instructed by the authors here

2. Download and install the implementations of the baseline methods reported in the manuscript. We recommend each method to be installed in a unique virtual environment in order to ensure that the correct dependencies are used.

3. Download the datasets used in the experiments:

   • StudentDB
   • Facebook
   • BlogCatalog
   • Flickr
   • YouTube
   • GR-QC
   • DBLP
   • PPI
   • Wikipedia

4. Modify the .ini configuration files from this folder to match the paths where the datasets are stored on your system as well as the paths where the methods are installed. Run the evaluation as:

   python -m evalne ./experiments/expLP1.ini

NOTE: In order to obtain the results for, e.g. different values of the embedding dimensionality, the conf file expLP1.ini has to be modified accordingly and the previous command rerun.

NOTE: For GAE/VGAE, AROPE, VERSE and the GEM library, special main.py files are required in order to run the evaluation through EvalNE. Once these methods are installed, the corresponding main file has to be added to the root folder of the method and called from the .ini configuration file. These main.py files are located in a ./main_files folder. For GIN and GatedGCN we directly provide the implementations used and main file under ./main_files