Deepwalk - Clone

Clone implementation of deepwalk paper in PyTorch.

• Source code by the original author

Install

Requirements:

• python >= 3.9

Commmand:

pip install git+https://github.com/helloybz/deepwalk-clone.git

Usage

deepwalk --data_root PATH --output_root PATH --config_file PATH [--gpu] [--checkpoint_period=0 INT]

--data_root is supposed to be a directory which contains at least 2 files, V and E.
V contains list of node ids separated by \n.
E contains list of pairs of node id, separated by \n.
The pair of node should has a form like [source_node_id]\t[target_node_id.

Example

data_root/V

1
2
3
4
5

data_root/E

1\t2
1\t4
2\t5
3\t1

Config file

All of the hyper-parameters of deepwalk is controlled in this yaml file.
Below is an example of the config file.

n_dims: 128 # referred as d in the paper.
random_walker:
  steps_per_walk: 40 # referred as t in the paper.
  walks_per_node: 80 # referred as r in the paper.
skipgram:
  lr: 0.025
  window_size: 10 # referred as w in the paper.

Experiments

CORA dataset, a citation network, is used.
Identical hyper parameters introduced in the paper are used to both the original Deepwalk and the cloned Deepwalk.\

• 128 dimensions
• 40 steps per walk
• 80 walks per node
• 10 window size (10+10+1 for both directions.)

Loss graph

Multi-Label Classification

• Logistic regression is used as classifier.
• Train:Test ratio varies from 1:9 to 9:1.
• Metrics are averaged after 10 runs.
• There are 7 classes in CORA dataset.

Micro F1

Column name means the percentage of train split.

Method	10%	20%	30%	40%	50%	60%	70%	80%	90%
Deepwalk-Clone	0.219	0.242	0.282	0.335	0.347	0.342	0.360	0.377	0.352
Deepwalk	0.200	0.206	0.215	0.223	0.213	0.236	0.242	0.256	0.247

Macro F1

Column name means the percentage of train split.

Method	10%	20%	30%	40%	50%	60%	70%	80%	90%
Deepwalk-Clone	0.411	0.421	0.427	0.457	0.478	0.468	0.477	0.511	0.454
Deepwalk	0.135	0.138	0.144	0.132	0.122	0.136	0.128	0.125	0.108