Early classification of time series based on uncertainty measure

The early classification of time series data is a critical problem in many time-sensitive applications such as health informatics. Where the prediction of class value, as early as possible, is highly valuable while preserving the accuracy as on full-length sequence data. For example, early diagnosis can provide better treatment to the patient or even save their lives. The aim of early classification is to analyse the sequence data at each time point continuously and predict the class label when a sufficient amount of data is available. Thus, the decision of early classification is a challenging task that needs to be addressed. Therefore, in this work, we propose an early classification model which relies on a set of probabilistic classifier and a confidence threshold that is measured in term of uncertainty. Formally, our model is divided into two parts. i) Learning phase, define the safeguard point for each class so that it makes sense to predict the label of any sequence with some acceptable accuracy. These safeguard points are identified based on user-defined accuracy. ii) Prediction phase, classify the time series only if the uncertainty of probabilistic output lie under the confidence threshold, that is obtained in the learning phase. We have evaluated our proposed model for 15 UCR datasets and compared with baseline state-of-art methods. Results clearly show that our proposed model is sianificantlv better in term of early classification.