Benchmarking Imputation Methods for Fuzzy Datasets

Imputation methods are widely used to replace missing values in datasets, thereby improving the overall quality of samples and enabling further statistical procedures. Various measures and tools have been proposed to compare the effectiveness and results of imputation algorithms. This paper describes the extended benchmarking approach designed explicitly for imputing fuzzy datasets. It is intended as a unique combination of classical tools with new measures that address the special features of fuzzy sets. With the help of this benchmark, five imputation methods (the widely known missForest, miceRanger, kNN, and PMM algorithms, as well as the dimp method aimed specifically for fuzzy data) are numerically compared using various synthetic, real-life, single- and multivariate datasets. It is the first such comparison explicitly related to fuzzy data. The obtained conclusions shed new light on the existing, yet still overlooked, problem of imputing missing fuzzy data.