Comparison of Measures for Characterizing the Difficulty of Time Series Classification

The performance of machine learning algorithms is influenced both by their characteristics and parameterization as well as by the properties of the data they are trained and evaluated on. The latter aspect is often neglected. In this paper, we focus our attention on properties of the data that affect the accuracy of time series classification. We experimentally study how the difficulty of classifying time series is related to well-known model-agnostic data complexity measures. Our experiments show that (a) many of these measures are highly correlated with classification scores such as accuracy and F1 and (b) different families of complexity measures capture different properties of the data.