Extração e seleção de características para a classificação eficiente de séries temporais

Abstract

As the production of time series increases, so does the need to mine them. Currently one of the most prominent mining tasks has been the time series classification. This task received many publications and solutions mainly focused on classification accuracy. This led to a state of the art specialized in high accuracy results, but also with a high processing time. This characteristic makes the solution usability infeasible for large scale problems. Aiming to obtain both accurate and fast results, this work proposes 4T. It is a dictionary-based algorithm of feature extraction and selection focused on the efficiency of time series classification. The efficiency was proposed in this dissertation as an evaluation metric and was defined as the fraction between score and fitting time of a classification. The results obtained by 4T show an average efficiency higher than the efficiency of the available state-of-the-art results. These results include two scores: accuracy and AUROC. Along with fitting time the scores were calculated by classifying 71 datasets of the UEA & UCR archive.