Individualized Error Estimation for Classification and Regression Models

Abstract

Estimating the error of classification and regression models is one of the most crucial tasks in machine learning. While the global error is capable to measure the quality of a model, local error estimates are even more interesting: on the one hand they contribute to better understanding of prediction models (where does and where does not work the model well), on the other hand they may provide powerful means to build successful ensembles that select for each region the most appropriate model(s). In this paper we introduce an extremely localized error estimation, called individualized error estimation (IEE), that estimates the error of a prediction model M for each instance x individually. To solve the problem of individualized error estimation, we apply a meta model ${M}^{{_\ast}}$. We systematically investigate various combinations of elementary models M and meta models M ^∗ on publicly available real-world data sets. Further, we illustrate the power of IEE in the context of time series classification: on 35 publicly available real-world time series data sets, we show that IEE is capable to enhance state-of-the art time series classification methods.