Normalization and Bias in Time Series Data

Abstract

Data normalization is an important preprocessing step in data mining and Machine Learning (ML) technique. Finding an acceptable approach to deal with time series normalization, on the other hand, is not an easy process. This is because most standard normalizing approaches rely on assumptions that aren’t true for the vast majority of time series. The first is that all time series are stationary, which means that their statistical characteristics, such as mean and standard deviation, do not vary over time. The time series volatility is assumed to be uniform in the second assumption. These concerns are not addressed by any of the approaches currently accessible in the literature. This research provides theoretical and experimental evidence, that normalizing time series data, can prove to be of utmost value by trimming non necessary data points and achieving minimum information loss, by using the concept of Minimal Time Series Representation (MTSR).