node2vec

This repository provides an implementation of node2vec extended with restart probabilities and ensembles:
The extensions are added by Koen Bouwman and Jerry Schonenberg
node2vec is introduced by Aditya Grover and Jure Leskovec.

Basic Usage

Example

To run node2vec on the email-Eu-core dataset, execute the following command from the project home directory:
python src/main.py --input email-Eu-core.edgelist --labels email-Eu-core.labels --output results-email-Eu-core

Options

You can check out the other options available to use with node2vec using:
python src/main.py --help

Options added with the added functionality

We have added the following parameters to configure the added functionality:

• To configure the bayesian optimisation:
  • --train_set to specify the proportion of dataset used for optimisation
  • --bayesian_opt to toggle Enable bayesian optimisation
  • --iter_bayesian to specify the number of iterations for bayesian optimisation
  • --scoring to specify how to evaluate each iteration of bayesian optimisation
  • --cross_validation to specify the size of cross validation
  • --replications to specify the number of replications to evaluate hyperparameter configuration
• To configure the restart method:
  • --restarts to toggle the restart functionality
  • --tau to set the $tau$ parameter
  • --omega to set the $\omega$ parameter
  • --epsilon to set the $\varepsilon$ parameter
  • --s to set the $s$ parameter
• To configure the ensemble method:
  • --partitions to define how many ensembles you want
  • --p now also supports a sequence of floats
  • --q now also supports a sequence of floats

post processing

To find the $\lambda$ and/or $p,q$-lists to use for partitions you can use post_processing.py

Example post processing

To run post_process on the email-Eu-core dataset, execute the following command from the project home directory:
python src/post_process.py --dir results-email-Eu-core --partitions 4 --read --write To run learn about the options for post_process execute the following command from the project home directory:
python src/post_process.py --help

Input

The supported input format is an edgelist:

node1_id_int node2_id_int <weight_float, optional>

The graph is assumed to be undirected and unweighted by default. These options can be changed by setting the appropriate flags.

Output

The output file directory contains the following

• cl_args.json: a file with the settings of all the calleble arguments
• The eval directory, which contains the following:
  • directories for each replication with an embeddings.pkl file containing the vector embedding of the input graph for that replication
  • results.csv: a file with the results of the classifier over all replications
  • best_settings.json: a file that contains the best settings for each calleble argument
• If the program was called with the --bayesian_opt flag the following will also be in the output directory:
  • BO_opt*.pdf: a plot of the bayesian optimisation
  • opt_results.pkl: the scores of each configuration of the bayesian optimisation run

Citing

If you find node2vec useful for your research, please consider citing the following paper:

@inproceedings{node2vec-kdd2016,
author = {Grover, Aditya and Leskovec, Jure},
 title = {node2vec: Scalable Feature Learning for Networks},
 booktitle = {Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining},
 year = {2016}
}