Characterising nuclear-norm matrix completion for nutrient imputation in food composition databases
Array · 2026-12 · 32 · 101428
DOI: 10.1016/j.array.2026.101428 ↗At a glance
Journal article in Array, 2026-12. Written by Fatma Indriani and 4 co-authors. 44 references. The full text is freely available (open access, CC BY 4.0).
Abstract (original language)
Missing nutrient values are a persistent challenge in national food composition databases (FCDBs), where costly laboratory analysis leaves many entries unobserved. This study characterises nuclear-norm matrix completion (SoftImpute) for nutrient imputation, evaluating its operating regime, mechanism sensitivity, downstream behaviour, computational cost, and identifiability on two FCDBs, Indonesia’s TKPI (1146 foods, 19 nutrients) and USDA SR Legacy (7793 foods, 18 nutrients). A 10-seed sweep masking 5% to 50% of observed entries on TKPI reveals a regime-dependent transition. KNN is significantly more accurate at low additional masking (5%–10%), no statistically detectable difference appears between 15% and 25%, and SoftImpute achieves significantly lower NRMSE from 30% onward. The same crossover recurs on USDA when values are missing completely at random or missing at random, but is obscured under missing-not-at-random censoring by a denominator artefact in standard NRMSE. PICA, a modern multivariate baseline, numerically matched or outperformed SoftImpute across much of the higher-masking regime on both databases despite non-normal nutrient distributions. Downstream classification and regression were evaluated transductively, completing each naturally incomplete matrix before cross-validation, and therefore reflect the utility of an already completed database, not generalisation to unseen foods. Here the best method depends on the objective, with simple imputers (mean, median) competitive for food-category classification, MICE achieving the highest mean regression scores on both datasets, and the tested masked-NMF implementation ranking last on all downstream tasks. Cell-wise reconstruction error alone is insufficient for method selection; practitioners should consider both the missingness regime and the intended analytical purpose.
Cite
APA
Indriani, F., Kartini, D., Budiman, I., Annisa, S., & Mahmudah, K. R. (2026). Characterising nuclear-norm matrix completion for nutrient imputation in food composition databases. Array, 32, 101428. https://doi.org/10.1016/j.array.2026.101428
Vancouver
Indriani F, Kartini D, Budiman I, Annisa S, Mahmudah KR. Characterising nuclear-norm matrix completion for nutrient imputation in food composition databases. Array. 2026;32:101428. doi:10.1016/j.array.2026.101428
BibTeX
@article{Indriani2026,
author = {Indriani, Fatma and Kartini, Dwi and Budiman, Irwan and Annisa, Selvi and Mahmudah, Kunti Robiatul},
title = {Characterising nuclear-norm matrix completion for nutrient imputation in food composition databases},
journal = {Array},
year = {2026},
volume = {32},
pages = {101428},
doi = {10.1016/j.array.2026.101428}
}
EUVIMED