bespoke-sscd

A Re-implementation of BeSpoke (Bakshi et al. ICDM'18) for semi-supervised community detection.

• Faster Training due to the vectorized implementation
• Faster Evaluation

Requirements

• python3
• numpy
• scipy
• scikit-learn
• tqdm

Usage

python run.py --dataset dblp --train_size 100 --n_roles 4 --n_patterns 5 --pred_size 50000

Outputs:

= = = = = = = = = = = = = = = = = = = =
##  Starting Time: 2019-12-05 14:00:32
[dblp] # Nodes: 317080 # TrainComms: 100 # TestComms: 4900
ComputeJaccards: 100%|███████████████████████████████████████████████████████| 317080/317080 [00:12<00:00, 25267.88it/s]
ExtractPercentiles: 100%|█████████████████████████████████████████████████████| 317080/317080 [00:33<00:00, 9365.59it/s]
NodeLocalFeature: 100%|██████████████████████████████████████████████████████| 317080/317080 [00:12<00:00, 25065.49it/s]
-> (All)  # Comms: 50000
AvgOverAxis0: Precision:0.56 Recall:0.32 F1:0.32 Jaccard:0.27
AvgOverAxis1: Precision:0.76 Recall:0.83 F1:0.75 Jaccard:0.68
AvgGlobal: Precision:0.66 Recall:0.57 F1:0.54 Jaccard:0.47
## Finishing Time: 2019-12-05 14:02:36
= = = = = = = = = = = = = = = = = = = =

(Extracting and evaluating 50K communities only took 2 minutes on my laptop!)

Check all available arguments:

python run.py --help

References

1. Bakshi et al. 2018. Semi-Supervised Community Detection Using Structure and Size. ICDM.
2. The official implementation by the authors: [repo].