PRUNE-pytorch

A pytorch implementation of PRUNE : Preserving Proximity and Global Ranking for Network Embedding

Desciprtion

PRUNE is an unsupervised generative approach for network embedding.

Design properties PRUNE satisfies: scalability, asymmetry, unity and simplicity.

The approach entails a multi-task Siamese neural network to connect embeddings and our objective, preserving global node ranking and local proximity of nodes.

Deeper analysis for the proposed architecture and objective can be found in the paper (please see - PRUNE)

Requirement

pytorch==0.4.0
python==3.5.2

Usage

Clone the repository

git clone https://github.com/j40903272/PRUNE-pytorch

Prepare data

Prepare a graph with edge lists in <from_node, to_node> format

Examples in : edgelist.txt

Training

python3 train.py ../example/edgelist.txt

The PRUNE model would be stored in prune.pt

The embedding weights would be in prune_weight.pkl

Example

import numpy as np
from preprocess import preprocess
from model import PRUNE
from train import train

graph = np.loadtxt("../example/edgelist.txt").astype(np.int64)
nodeCount = int(graph.max()) + 1
data_loader, PMI_dict = preprocess(graph)

model = PRUNE(nodeCount).cuda()
train(model, data_loader)

emb_weight = model.node_emb.weight.data.cpu().numpy()
with open('prune_weight.pkl', 'wb') as f:
    pickle.dump(emb_weight, f)