A PyTorch + torch-geometric implementation of LSS, as described in the paper: Kangfei Zhao, Jeffrey Xu Yu, Hao Zhang, Qiyan Li, Yu Rong. A Learned Sketch for Subgraph Counting
python 3.7
networkx
numpy
scipy
scikit-learn
torch >= 1.5.0
torch-geometric 1.6
Import the conda environment by running
conda env create -f gpcard.yml
conda activate gpcard
Running cross validation
python active_train.py \
--dataset aids \
--embed_type prone \
--mode cross_val
Running active learning
python active_train.py \
--dataset aids \
--mode train \
--embed_type prone \
--active_iters 1 \
--uncertainty consist
Running ensemble (active) learning
python active_train.py \
--dataset aids \
--mode ensemble \
--embed_type prone \
--ensemble_num 5 \
--active_iters 1
All the parameters with their default value are in active_train.py
| name | type | description |
|---|---|---|
| model_type | String | type of GCN layer (GIN, GINE, GAT, NN, GCN, SAGE, NNGIN, NNGINConcat) |
| num_layers | Int | number of GNN layers |
| embed_type | String | the node feature encoding, (freq, n2v, prone, nrp, prone_concat) |
| pool_type | String | type of the substructure pooling layer (att, mean, max, sum) |
| epochs | Int | number of training epochs |
| batch_size | Int | mini-batch size for sgd |
| num_fold | Int | number of folder for cross validate |
| k | Int | number of hops to decompose query graphs |
| learning rate | Float | learning rate |
| multi_task | Bool | enable/disable the count classification task, for active learning |
| num_ensemble | Int | number of LSS modes for model ensemble |
| active_iters | Int | number of iteration for active learning |
| budget | Int | number of query to selected for each iteration of active learning |
| uncertainty | String | uncertainty function for active learning, e.g., entropy, margin, confident |
active_train.py # project extrance, parameters settings
train.py # deprecated codes for only supervied training
QuerySampler.py # data and query graph preprocessing, query decomposition
Queryset.py # input encoding, torch dataloader
util.py # model load/save
baseline.py # old codes for result evaluation
/active
active_util.py
ActiveLearner.py # source of the active learner
/cardnet # source of the LSS model
GINlayers.py # revised GIN layers (NNGIN, NNGINConcat)
layers.py # basic layers
model.py # LSS model
/scripts # old scripts for testing
/data # aids dataset sample to run the code
/dataset/aids # graph data
/queryset_homo/aids # queries
/true_homo/aids # true count of the queries in queryset_homo
/prone/aids # prone embdding matrix and node mapping file
To use your own dataset, you can put the data graphs, queries, true counts, and embedding to '/data/dataset/DATASET_NAME', '/data/queryset_homo/DATASET_NAME', '/data/true_homo/DATASET_NAME', '/data/prone/DATASET_NAME', respectively, and set the args.dataset to DATASET_NAME in 'active_train'.py
The format of input query graph and data graph follows GCARE.
Open an issue or send email to zkf1105@gmail.com if you have any problem
@inproceedings{DBLP:conf/sigmod/ZhaoYZLR21,
author = {Kangfei Zhao and
Jeffrey Xu Yu and
Hao Zhang and
Qiyan Li and
Yu Rong},
title = {A Learned Sketch for Subgraph Counting},
booktitle = {{SIGMOD} '21},
pages = {2142--2155},
publisher = {{ACM}},
year = {2021}
}