A Complexity-Invariant Measure Based on Fractal Dimension for Time Series Classification

Abstract

Classification is an important task in time series mining. It is often reported in the literature that nearest neighbor classifiers perform quite well in time series classification, especially if the distance measure properly deals with invariances required by the domain. Complexity invariance was recently introduced, aiming to compensate from a bias towards classes with simple time series representatives in nearest neighbor classification. To this end, a complexity correcting factor based on the ratio of the more complex to the simpler series was proposed. The original formulation uses the length of the rectified time series to estimate its complexity. In this paper the authors investigate an alternative complexity estimate, based on fractal dimension. Results show that this alternative is very competitive with the original proposal, and has a broader application as it does neither depend on the number of points in the series nor on a previous normalization. Furthermore, these results also verify, using a different formulation, the validity of complexity invariance in time series classification.