Automatic selection of time series imputation algorithms, DAGStat Conference 2019, Munich, Germany

Poster 'Automatic selection of time series imputation algorithms' at DAGStat Conference 2019 in Munich, Germany.

Abstract:

Replacement of missing values (imputation) in univariate time series is an often-used preprocessing step before further analysis. Hereby it is not uncommon, that multiple hundreds of time series have to be imputed at once — making an individual selection of the right imputation algorithm unfeasible. Additionally, even when just a few imputed time series are required, there might be nobody suitably trained or willing to spend time to select an appropriate imputation algorithm (which might partly explain the still widespread usage of rather simplistic time series imputation methods).

In all these circumstances, automatic imputation can be an essential tool. Automatic imputation hereby has to select the most suitable time series imputation algorithm, estimate the parameters and compute the imputed values.

In this poster, we take a deeper look at automatic imputation especially for univariate, equi-spaced time series. We explore to what extent approaches from classical imputation literature can be applied to univariate time series. We propose promising adaptions of algorithms and afterwards compare the results on a limited number of time series. Special focus hereby lies on evaluating if the proposed methods can deal with time series features like seasonality and trend.

KEYWORDS: Imputation, Interpolation, Time Series, Time Series Imputation