Interpolation-Prediction Networks

This repository is an official implementation of Interpolation-Prediction Networks for Irregularly Sampled Time Series, accepted at ICLR 2019. In this work, we present a new deep learning architecture for addressing the problem of supervised learning with sparse and irregularly sampled multivariate time series.

Requirements

The code requires Python 3.7 or later. The file requirements.txt contains the full list of required Python modules.

pip install -r requirements.txt

Training and Evaluation

For running our model on univariate time series (UWave dataset):

python src/univariate_example.py --epochs 1000 --hidden_units 2048 --ref_points 128 --batch_size 2048

To reproduce the results on MIMIC-III Dataset, first you need to have an access to the dataset which can be requested here. Once your application to access MIMIC has been approved, you can download the data. MIMIC is provided as a collection of comma-separated (CSV) files. You can use these scripts to import the csv files into a database. Assuming you installed postgres while creating the database, you need to install psycopg2 using

pip install psycopg2

Once the database has been created, run these scripts in order.

python src/mimic_data_extraction.py
python src/multivariate_example.py --epochs 1000 --reference_points 192 --hours_from_adm 48 --batch_size 256 --gpus 4

Results

Classification and regression performance on MIMIC-III.

Data Format Example

The notations here align with the notation section 3.1 in the paper. For brevity, lets assume we have just one example in the training set and dimension d = 2.

Reference

@inproceedings{
shukla2018interpolationprediction,
title={Interpolation-Prediction Networks for Irregularly Sampled Time Series},
author={Satya Narayan Shukla and Benjamin Marlin},
booktitle={International Conference on Learning Representations},
year={2019},
url={https://openreview.net/forum?id=r1efr3C9Ym},
}

Contact

For more details, please contact [email protected].