Imputation of Missing Data in Materials Science through\nNearest Neighbors and Iterative Predictions

Missing data in tabular data sets\nis ubiquitous in statistical\nanalysis, big data analysis, and machine learning studies. Many strategies\nhave been proposed to impute missing data, but their reliability has\nnot been stringently assessed in materials science. Here, we carried\nout a benchmark test for six imputation strategies: Mean, MissForest,\nHyperImpute, Gain, Sinkhorn, and a newly proposed MatImpute on seven\nrepresentative data sets in materials science. The imputation-induced\nerrors (IIEs) were evaluated through the difference between imputed\nand original values, by root mean square error (RMSE), Wasserstein\ndistance (WD), and a newly introduced metrics data set correlation\nconvergence (DCC), to measure the difference at three aspects for\nindividual data, column-wise distribution, and correlation stability\nof a data set. MatImpute outperformed the others with the least RMSE\nand WD and the highest DCC. The IIE increases with the increase of\ndata missing ratio and in the order of missing at random < missing\ncompletely at random ≤ missing not at random, considering inherent\ncorrelations among missing data. A similar trend was observed for\nthe increase of IIE along the central departure distance in units\nof the standard deviation, which is consistent with the increase of\ndifficulty from interpolation to extrapolation. Further tests of IIE\nin regression and classification machine learning predictive models,\nMatImpute also preserved the highest data recovery fidelity. We released\nthe code of MatImpute to facilitate the construction of high-quality\ndata sets in materials science.