Interpolation-Prediction Networks

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. The architecture is based on the use of a semi-parametric interpolation network followed by the application of a prediction network. The interpolation network allows for information to be shared across multiple dimensions of a multivariate time series during the interpolation stage, while any standard deep learning model can be used for the prediction network.

We use a two layer interpolation network. The first interpolation layer performs a semi-parametric univariate interpolation for each of the D time series separately while the second layer merges information from across all of the D time series at each reference time point by taking into account the correlations among the time series.

Reference

Satya Narayan Shukla and Benjamin Marlin. Interpolation-prediction networks for irregularly sampled time series. In International Conference on Learning Representations, 2019. [pdf]

Requirements

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

Usage

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

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.

Contact

For more details, please contact [email protected].