RBQR

Prerequisites

• Linux or macOS
• Python 2 or 3
• scipy
• sklearn

Dataset

These datasets are public datasets.

• PPI contains 3,890 nodes, 76,584 edges and 60 labels.
• Wikipedia contains 4,777 nodes, 184,812 edges and 40 labels.
• Blogcatalog contains 10,312 nodes, 333,983 edges and 39 labels.
• flickr contains 80,513 nodes, 5,899,882 edges and 195 labels.
• Youtube contains 1,138,499 nodes, 2,990,443 edges and 47 labels.

Training

Training on the existing datasets

Create emb directory to save output embedding file

mkdir emb

You can use python RBQR_NE.py -graph example_graph to train RBQR model on the example data.

If you want to train on the dataset, you can run

python RBQR_NE.py -graph data/PPI.mat -emb1 emb/PPI_sparse.emb -emb2 emb/PPI_spectral.emb -dimension 128
python RBQR_NE.py -graph data/POS.mat -emb1 emb/POS_sparse.emb -emb2 emb/POS_spectral.emb -dimension 128
python RBQR_NE.py -graph data/blogcatalog.mat -emb1 emb/blogcatalog_sparse.emb -emb2 emb/blogcatalog_spectral.emb -dimension 128
python RBQR_NE.py -graph data/flickr.mat -emb1 emb/flickr_sparse.emb -emb2 emb/flickr_spectral.emb -dimension 128
python RBQR_NE.py -graph data/youtube.mat -emb1 emb/youtube_sparse.emb -emb2 emb/youtube_spectral.emb -dimension 128

Where PPI_sparse.emb and PPI_spectral.emb are output embedding files and dimension=128 is the default setting for a good result. If you want to evaluate the embedding without spectral filter via node classification task, you can run

python predict.py --label data/blogcatalog.mat --embedding emb/blogcatalog_sparse.emb.npy --start-train-ratio 10 --stop-train-ratio 90 --seed 0
python predict.py --label data/POS.mat --embedding emb/POS_sparse.emb.npy --start-train-ratio 10 --stop-train-ratio 90 --seed 0
python predict.py --label data/PPI.mat --embedding emb/PPI_sparse.emb.npy --start-train-ratio 10 --stop-train-ratio 90 --seed 0
python predict.py --label data/flickr.mat --embedding emb/flickr_sparse.emb.npy --start-train-ratio 1 --stop-train-ratio 9 --seed 0
python predict.py --label data/youtube.mat --embedding emb/youtube_sparse.emb.npy --start-train-ratio 1 --stop-train-ratio 9 --seed 0

If you want to evaluate the embedding with spectral filter via node classification task, you can run

python predict.py --label data/blogcatalog.mat --embedding emb/blogcatalog_spectral.emb.npy --start-train-ratio 10 --stop-train-ratio 90 --seed 0
python predict.py --label data/POS.mat --embedding emb/POS_spectral.emb.npy --start-train-ratio 10 --stop-train-ratio 90 --seed 0
python predict.py --label data/PPI.mat --embedding emb/PPI_spectral.emb.npy --start-train-ratio 10 --stop-train-ratio 90 --seed 0
python predict.py --label data/flickr.mat --embedding emb/flickr_spectral.emb.npy --start-train-ratio 1 --stop-train-ratio 9 --seed 0
python predict.py --label data/youtube.mat --embedding emb/youtube_spectral.emb.npy --start-train-ratio 1 --stop-train-ratio 9 --seed 0