Time Series Qlet: Invariant approach for data mining

Abstract

In the last decade or so the problem of Time Series Classification (TSC) has attained considerable research interests in both the Machine learning and Datamining community. In a TSC problem any real valued data is considered as a time series and the challenge is to find the best discriminating features based on the ordering of variables. Recently, the research interest has shifted to Time Series Shapelets which are small local patterns in a time series that are highly predictive of a class and are thus very useful features for building classifiers. Most of the work in time-series shapelet till date makes use of a supervised learning approach. One of the major drawbacks with the supervised shapelet approach is that they consider global features of the time series dataset to compute a shapelet which in turn leads to more time and space complexity issues. We believe that instead of taking global features of a time series data only the local features that are of interest to the problem at hand should be considered. How to find these local features from a given time series data set and to get the natural shaplet by an unsupervised approach is an interesting problem as such and in this paper we aim to propose solutions to this end. Here we introduce a new approach for time series, called Time Series Qlet, a novel approach for data mining, which addresses the above limitations. Qlets are quantized continuous subsequences of data obtained after a natural break (chunking) of quantized time series data set. Moreover we focus only on local features of a data-set ignoring the global features.