graph2vec v2

graph2vec is to node2vec as doc2vec is to word2vec

node2vec performs random walks on a graph in order to create sequences that can be fed into word2vec graph2vec performs random walks on multiple graphs and feeds those walks, along with a unique graph ID per graph into doc2vec, which learns a continuous embedding for each node and each graph.

Note that this technique is the same as this graph2vec but they use a more complicated "word" equivalent--rooted subgraphs around each node--rather than the random walks that are performed here. And the results here are better than the ones reported there.

Basic Usage

Example

First preprocess the datasets:

python src/process_graph.py graph/PROTEINS

This separates the graphs into individual gpickled networkx files located in the DATASET/graphs/ folder

Then run graph2vec to extract the embeddings:

python src/graph2vec.py graph/PROTEINS/graphs --p 1 --q 0.5 --dimensions 1024 --output emb/PROTEINS_p1_q.5_d1024.emb

This will save the embedding

Then run the classification

python src/classify.py graph/PROTEINS

which will output the accuracy on a random test set of 10% of the data. An SVM is used to train the data, with parameter tuning done using 5-fold cross validation on the train set.

Results

Dataset	Score
PROTEINS	0.836 (+/- 0.06)
MUTAG	0.842 (+/- 0.18)
NCI1	0.808 (+/- 0.03)
NCI109	0.785 (+/- 0.04)
PTC	0.697 (+/- 0.12)