A Large Comparison of Normalization Methods on Time Series

Normalization is a mandatory preprocessing step in time series problems to guarantee similarity comparisons invariant to unexpected distortions in amplitude and offset. Such distortions are usual for most time series data. A typical example is gait recognition by motion collected on subjects with varying body height and width. To rescale the data for the same range of values, the vast majority of researchers consider z-normalization as the default method for any domain application, data, or task. This choice is made without a searching process as occurs to set the parameters of an algorithm or without any experimental evidence in the literature considering a variety of scenarios to support this decision. To address this gap, we evaluate the impact of different normalization methods on time series data. Our analysis is based on an extensive experimental comparison on classification problems involving 10 normalization methods, 3 state-of-the-art classifiers, and 38 benchmark datasets. We consider the classification task due to the simplicity of the experimental settings and well-defined metrics. However, our findings can be extrapolated for other time series mining tasks, such as forecasting or clustering. Based on our results, we suggest to evaluate the maximum absolute scale as an alternative to z-normalization. Besides being time efficient, this alternative shows promising results for similarity-based methods using Euclidean distance. For deep learning, mean normalization could be considered.