Early classification of time series

An increasing number of applications require to recognize the class an incoming time series as quickly possible without unduly compromising accuracy prediction. In this paper, we put forward a new optimization criterion which takes into account both cost misclassification and delaying decision. Based on criterion, derived family non-myopic algorithms try anticipate expected future gain in information balance with waiting. one algorithms, unsupervised-based, expectations use clustering series, while second class, supervised-based, are grouped according confidence level classifier used label them. Extensive experiments carried out real datasets using large range delay functions show that presented able solve earliness vs. trade-off, supervised partition based approaches faring better than unsupervised ones. addition, all these methods perform wide variety conditions state art method myopic strategy is recognized being very competitive. Furthermore, our feature proposed explains part obtained performances.