Concatenated Decision Paths Classification For Time Series Shapelets

Abstract

Time-series classification is widely used approach for classification. Recent development known as time-series shapelets, based on local patterns from the time-series, shows potential as highly predictive and accurate method for data mining. On the other hand, the slow training time remains an acute problem of this method. In recent years there was a significant improvement of training time performance, reducing the training time in several orders of magnitude. Reducing the training time degrade the accuracy in general. This work applies combined classifiers to achieve high accuracies, maintaining low training times-in the range from several second to several minutes-for datasets from the popular UCR database. The goal is achieved by training small 2,3-nodes decision trees and combining their decisions in pattern that uniquely identifies incoming time-series.