Generating synthetic time series to augment sparse datasets

This is the supporting web page for our paper titled "Generating synthetic time series to augment sparse datasets".

This paper has been accepted for publication to the ICDM 2017.

Example of synthetic time-series:

Example of a generated synthetic sample (right) for Cylinder class of CBF dataset (left) by averaging a set of time-series taken from the class:

      

Experimental results

We used the 85 datasets of UCR Archive.

The detailed results where we ranked each synthetic sample generation method with 2 to 6 samples available per class are available here.

All the plots for the experiments where we progressively augmented the number of samples available are available here.

The detailed results where we doubled the size of the full training with synthethic sample are available here.

The source code

The software is developed in Java and the source code of our method can be downloaded here : download the source code (TimeSeriesGenerator.java)

It requires The Apache Commons Mathematics Library.

We will release it open-source upon acceptance of the paper.


Last update : May 2017