PNC can be accessed from NIH, and ABCD can be accessed from NIMH Data Archive.
For those who can not access ABCD and PNC datasets, we also provide an open-source dataset, ABIDE. Please follow the instruction to download and process this dataset.
python main.py --config_filename setting/pnc_fbnetgen.yamlpython main.py --config_filename setting/abcd_fbnetgen.yamlpython main.py --config_filename setting/abide_fbnetgen.yamlAll hyperparameters can be tuned in setting files.
model:
# For the model type, there are 3 choices: "seq", "gnn" or "fbnetgen".
type: fbnetgen
# For the feature extractor, there are two choices: "gru" or "cnn".
extractor_type: gru
# For the feature extractor, there are two choices: "product" or "linear".
# We suggest using "product" since it is faster.
graph_generation: product
# Two hyperparameters are tuned in our paper.
embedding_size: 8
window_size: 8
train:
# For the training method, there are two choices: "normal" or "bilevel".
# "bilevel" will be in effect only if the model.type is set as "fbnetgen"
# We suggest using "normal".
method: normal
# If the model.type is set as "gnn", this hyperparameter will be in effect.
# There are 2 choices: "uniform" or "pearson".
pure_gnn_graph: pearson| method | Dataset | AUROC | Accuracy |
|---|---|---|---|
| FBNETGNN | PNC | 80.8±3.3 | 74.8±2.4 |
| FBNETGNN | ABCD | 94.5±0.7 | 87.2±1.2 |
| FBNETGNN | ABIDE | 72.74±4.26 | 66.31±3.71 |
Our model's performance is not as good as PNC in ABIDE since ABIDE is collected from different sites, making it heterogeneous and challenging to train a model.