Real-world time series often contain missing values due to sensor failures, power outages and transmission problems. The recovery of these missing values allows better analysis of time series. Several methods have been proposed to recover missing values in time series, which can be matrix-based, pattern-based or machine learning-based technique. Selecting “the best” recovery method highly depends on the dataset properties and often requires users to run multiple approaches with a different set of configuration parameters.

The aim of this thesis is to study and compare different ways to perform a recommendation of recovery techniques. The recommendation will relieve the user from the task of selecting and configuring recovery algorithms The thesis will focus on two classes: feature-based approach and parameter-based approach. The output of the thesis will be a solution that allows to select the most appropriate recovery technique in in a systematic way more…