MGMC - Multigraph Geometric Matrix Completion (MGMC)

1. Requirements

To make sure you have the right packages.

pip install -r requirements.txt

2. Dataset

Make sure dataset in ./data/ folder. Alternatively, you can load your own dataset by implementing the load_data method after inheriting MultiGMCTrainer class from trainer.py. Just make sure you change the following variables below. Please check MultiGMCTrainerPPMI in trainer.py as an example.

class MultiGMCTrainerPPMI(MultiGMCTrainer):
    # ... 
    def load_data(self):
        # Just make sure you change the following variables accordingly.
        self.data_x = data_x
        self.data_y = data_y
        self.data_meta = data_meta
        self.M = M
        self.initial_tr_mask = initial_tr_mask
        self.Lrow = Lrow
        self.A_matrix_row = A_matrix_row
        self.impute_idx = impute_idx

3. Train model

Now we are ready to train the model including hyperparameter optimization.

python main.py --train_hyperopt=1

4. Test model

Test the model given the hyperparameters from the previous step.

python main.py --train_hyperopt=0

5. Check output results

Output results will be saved at ./gmc_output/. You will find a pickle file containing a list of 4 elements (experiments from randomly dropping of entries. Each element is a list of outputs coming from the outputs from the ten-fold Stratified-CV.

# load pickle file
with open('./gmc_output/output.pkl', 'rb') as load_file:
    loaded_pickle_obj = pickle.load(load_file)

loaded_pickle_obj # this is now a list containing 4 lists
loaded_pickle_obj[0] # this is a list with 10 elements

loaded_pickle_obj[0][0] is a tuple with the following information:

(ground_truth_test_set, 
 predicted_probabilities_test_set, 
 predicted_labels_test_set, 
 RMSE)

Citation

If you use this code, please cite:

@misc{MGMC_2020,
    title={Simultaneous imputation and disease classification in incomplete medical datasets using Multigraph Geometric Matrix Completion (MGMC)},
    author={Gerome Vivar and Anees Kazi and Hendrik Burwinkel and Andreas Zwergal and Nassir Navab and Seyed-Ahmad Ahmadi},
    year={2020},
    eprint={2005.06935},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}