Blind Feature Extraction for Time-Series Classification Using Haar Wavelet Transform

Abstract

Time-series classification has attracted increasing interest in recent years, particularly for long time-series as those arising in bioinformatics and financial domain. Many dimensionality reduction algorithms have been proposed to attack the so-called curse of dimensionality problem. However, choosing the number of features is not a trivial task and has not been well considered. In this paper, we propose a novel blind feature extraction algorithm with Haar wavelet transform which can determine the feature dimensionality automatically. The algorithm takes the tradeoff of achieving lower dimensionality and lower sum of squared errors between the features and original time-series. Experimental results performed on several widely used time-series data demonstrate the effectiveness of the proposed algorithm.