Overview

BeatLex is an algorithm that succintly summarizes and forecasts time series data. It is designed for data containing patterns that occur repeatedly, especially if these patterns are complex and nonlinear, change over time, and may distortions in their shape or length.
BeatLex has the following properties:

• Fast and online: requires linear time in the data size, and bounded memory (it does not require the full history to process future data).
• Effective: it provides accurate pattern labels and forecasts.
• Principled and parameter-free: it is fit using the Minimum Description Length principle of summarizing data by compressing it into as few bits as possible.
• General: it applies to any type of time series data, and can use multidimensional (i.e. coevolving) time series.

Code

The source code and experiments used in the paper are available: [Download]

Datasets

The MITDB dataset used in the paper can be found at PhysioNet https://physionet.org/cgi-bin/atm/ATM?database=mitdb.

The MOCAP data can be found at http://mocap.cs.cmu.edu. The preprocessed time series as used in our paper can be found in our source code package.