GraphNodeClassification

This is the Python-3.6 re-implementation of the paper DeepWalk: Online Learning of Social Representations.

Overview

There are two different implementations within this repository. The deep walk re-implementation is located within deepwalk folder. Another implementation that uses negative sampling instead of hierarchical softmax is located in src. We run our experiments on the BlogCatalog dataset.

Dependencies

torch-1.2.0
torchvision-0.4.0
numpy-1.17.2
sklearn-0.21.3
matplotlib-3.0.2
tqdm
h5py
pickle

Deep Walk Re-implementation

The deepwalk re-implementation is included in the ./deepwalk directory.

$ cd deepwalk

To train deepwalk algorithm on BlogCatalog dataset, use

$ python3 train.py

Deepwalk can be trained on any other graph by passing its node list and edge connection files in the same format as BlogCatalog inside the main function. The final embeddings are stored into embeddings.pkl in the same folder. To change any arguments associated with deepwalk training such as random walk length, modify values present in args.py

$ python3 inference.py

To output f1 scores for the given embeddings file and train-test split ratio. The path to embeddings pkl file and the ratio need to be changed in inference.py before proceeding.

Negative sampling implementation

The source files for this implementation are located in ./src directory.

$ python3 train.py

By default, this script starts training on BlogCatalog dataset using the default parameters set in train.py. The generated embeddings are dumped into ./dump/embeddings.pkl.

$ python3 inference.py

This script loads the learnt embeddings from ./dump/embeddings.pkl and trains a linear SVM to report classification accuracy.

$ python3 visualise.py

This script shows the embedding separation between close nodes and far nodes in the graph.

Results

Re-implementation results

Percentage of labelled nodes	10%	20%	30%	40%	50%	60%	70%	80%	90%
Train Macro F1	33.7	31.3	29.54	29.51	29.94	28.22	30.04	27	30.7
Train Micro F1	40	38.7	36.8	37.12	36.6	35.43	36.16	34.83	36
Test Macro F1	17.1	20.11	23.55	23.71	24.77	26.68	26.07	28.09	28.15
Test Micro F1	28.3	29.8	31.9	31.83	31.98	34	32.94	35.72	33.45

Negative sampling results