Machine Learning-based Early Decision Making

The primary focus of this HDR manuscript is on the Early Classification of Time Series (ECTS) problem, that involves determining the optimal time to initiate the classification of an incoming time series observed over time, while balancing the trade-off between earliness and accuracy of the predictions. The following contributions to this field are presented in this document: (i) the release of an open-source library, accompanied by an extensive benchmark; (ii) the development of a non-myopic ECTS framework, called Economy; (iii) the adaptation of this framework to revocable decisions allowing decisions to be changed based on new available measurements; (iv) a generic method to adapt ECTS approaches to open time series, which have no predefined beginning or end and are observed as data streams. Looking ahead, an extensive research program is developed for the coming years. This research program extends the ECTS problem to a more generic one, called ML-EDM (Machine Learning based Early Decision Making), and proposes ten challenges to the scientific community to deal with a wide range of situations where data is collected over time. Three challenges have been presented in relation to the learning task at hand: extending ML-EDM to unsupervised learning (challenge \#1), to regression tasks (challenge \#2) and to weakly-supervised learning (challenge \#3). The development of data type agnostic ML-EDM approaches has been singled out as an important direction of research to extend the domains of application, yielding challenge \#4. Extending ML-EDM to the online scenario has also been identified as important too which raises three challenges (challenge \#5, \#6 and \#7). Being able to revoke decisions properly is significant as well in many applications and raises two challenges (\#8 and \#9). The origin of the different costs involved in the optimization of decision times has been discussed, leading to a last challenge (challenge \#10) to extend the ML-EDM problem to cases where these costs vary over time. Finally, a range of application areas for which ML-EMD could lead to significant progress in the near future have been described, such as anomaly detection, predictive maintenance, patient health monitoring, self-driving vehicles.