Add seed parameter.

[gerritjandebruin]

Hi Elior Cohen (and others),

This is my first pull request ever. Please help me a bit when I miss certain things.

A while ago we noticed that this Node2Vec implementation yields different results, even when run in one Python process. This problem is also mentioned by @marcovarrone in Issue #26.

In this pull request, we (@pereirabarataap and me) propose to add a simple parameter seed, which allows to set the random seed of both Numpy and the Python random module.

This issue fixes:

• When calling Node2Vec with the same parameters, will obtain the exact same embedding in one process. We show this in the another-example.py at https://github.com/gerritjandebruin/node2vec/blob/another-example/another-example.py.

Future work:

• Make the seed also work with multiple workers.
• Make the seed work between different Python process.
  At this moment it does not, also not with a fixed PYTHONHASHSEED.

Kind regards,

Gerrit-Jan and António

[eliorc]

If I understand you correctly - do you mean this works only for workers=1?

[gerritjandebruin]

Yes. This is also the case for gensim (https://radimrehurek.com/gensim/models/word2vec.html).

[eliorc]

Please view my comments

[eliorc]

Hi, please change the target of the request to the dev branch, I want to include some changes of my own @gerritjandebruin

[gerritjandebruin]

These two commits fix a uncorrelated typo and add additional explanation about the seed parameter.

The working of the seed parameter can be checked with the following example:

import random

import networkx as nx
from node2vec import Node2Vec
import numpy as np
import pandas as pd

random.seed(0)
np.random.seed(0)

G = nx.karate_club_graph()

def get_embedding():
    model = Node2Vec(G, dimensions=4, seed=0)
    embedding = model.fit()
    return pd.DataFrame(embedding.wv.vectors, index=embedding.wv.index2entity)

print('Embedding 1')
print(get_embedding().sort_index().head())

print('Embedding 2')
print(get_embedding().sort_index().head())