Codebase for "Estimating Missing Data in Temporal Data Streams Using Multi-Directional Recurrent Neural Networkss (MRNN)"

Authors: Jinsung Yoon, William R. Zame, Mihaela van der Schaar

Paper: Jinsung Yoon, William R. Zame, Mihaela van der Schaar, "Estimating Missing Data in Temporal Data Streams Using Multi-Directional Recurrent Neural Networks," IEEE Transactions on Biomedical Engineering, 2019.

Paper Link: https://ieeexplore.ieee.org/document/8485748

Contact: [email protected]

This directory contains implementations of MRNN framework for imputation in time-series data using GOOGLE stocks dataset.

To run the pipeline for training and evaluation on MRNN framwork, simply run python3 -m main_mrnn.py.

Command inputs:

• file_name: data file name
• seq_len: sequence length of time-series data
• miss_rate: probability of missing components (to be introduced)
• h_dim: hidden state dimensions
• batch_size: the number of samples in mini batch
• iteration: the number of iteration
• learning_rate: learning rate of model training
• metric_name: imputation performance metric

Example command

$ python3 main_mrnn.py --file_name data/google.csv --seq_len 7 
--missing_rate: 0.2 --h_dim 10 --batch_size 128 --iteration 2000
--learning_rate 0.01 --metric_name rmse

Outputs

• x: original data with missing
• ori_x: original data without missing
• m: mask matrix
• t: time matrix
• imputed_x: imputed data
• performance: imputation performance