Early Classification of Time Series: methods and applications

A whole series of problems involves identifying the class associated with an observation (for example, the type of tumor associated with a tomography scan). But the world is also made up of evolving processes, and therefore of problems requiring classification to take place while measurements relating to the phenomenon of interest are taking place over time. Forecasting a volcanic eruption is a case in point. The decision-maker is faced with a dilemma: the earlier the prediction of an imminent eruption, the easier it will be to take the necessary action, but the more uncertain the prediction, and a false alarm can be very damaging for the future. 

Waiting for additional measurements makes prediction more certain, but at the cost of making the action to be taken much more costly. Many less dramatic applications fall into this category. For example, monitoring a cold chain for foodstuffs involves detecting any deviations impacting on the quality of the goods as soon as possible, while avoiding falling victim to too many false alarms. Applications in agriculture, medicine, predictive maintenance and anomaly detection in general are numerous. 

How to optimize online the tradeoff between the earliness and the accuracy of the decision is the object of the early classification of time series problem (ECTS). 

This talk will start by questioning how to express the challenge of ECTS in formal terms. It will then present a few representative approaches and report empirical experiments about their performance and their ability to tackle a variety of situations.