Feature Selection and Missing-Data Imputation for ML Prediction of Phosphor Emission Wavelengths Evaluated by Reconstruction Accuracy Versus Predictive Utility

Missing values in materials datasets can compromise both data reconstruction and the prediction of material properties. However, an imputation method that accurately reconstructs individual descriptors does not necessarily maximise downstream predictive performance. Here, we investigate this distinction using 2,363 phosphor records from the Inorganic Phosphor Optical Properties database and compare three descriptor representations: a combined set of 88 features, a 13-feature physics-informed set, and a 75-feature data-driven set. Mean, K-nearest-neighbour, iterative Bayesian-regression and random-forest-based imputation methods are evaluated under artificially generated MCAR, MAR and MNAR mechanisms at missingness rates of 10–50%. Reconstruction error is assessed separately from emission-wavelength prediction using five regression algorithms and host-grouped cross-validation. In the evaluated configurations, random-forest regression performs best after feature-count selection, while the physics-informed representation approaches the predictive performance of the combined feature set. The iterative random-forest imputer achieves the lowest average normalised reconstruction error across all three representations. Nevertheless, the imputer minimising reconstruction error does not consistently maximise downstream predictive performance, with agreement between the two criteria in only 50.2% of the examined model-specific comparisons. These findings highlight the need to evaluate imputation according to its intended use rather than relying on reconstruction metrics alone. The reported performance remains conditional on the descriptor-selection procedure, host grouping and simulated missingness mechanisms; independent validation is required before extrapolation to chemically novel phosphors.