|
Hypervolume rankings decouple from unbiased predictive performance in multi-objective wavelength selection for NIR spectroscopy of Wolffia globosa |
|---|---|
| รหัสดีโอไอ | |
| Creator | Somboon Sukpancharoen |
| Title | Hypervolume rankings decouple from unbiased predictive performance in multi-objective wavelength selection for NIR spectroscopy of Wolffia globosa |
| Contributor | Nonpravit Chianrungsaeng |
| Publisher | Faculty of Engineering, Khon Kaen University |
| Publication Year | 2569 |
| Journal Title | Agricultural and Biological Engineering |
| Journal Vol. | 3 |
| Journal No. | 4 |
| Page no. | 109-121 |
| Keyword | Near-infrared spectroscopy, Wavelength selection, Multi-objective optimisation, Nested cross-validation, Selection bias |
| URL Website | https://ph04.tci-thaijo.org/index.php/abe |
| Website title | ThaiJo |
| ISSN | 3056-932X (Online) |
| Abstract | Multi-objective metaheuristics are widely used to select informative wavelengths in near-infrared (NIR) spectroscopy, and their quality is conventionally assessed with Pareto-based indicators such as hypervolume. Such indicators assume deterministic objectives, yet one objective here, the cross-validated prediction error, is an estimator carrying non-negligible variance, and the consequences of treating it as fixed remain largely unquantified. We benchmark four binary multi-objective metaheuristics (NSGA-II, MO-BPSO, MO-BDE, MO-bGWO) and the CARS selector on a 50-sample, 474-wavelength NIR protein calibration problem in Wolffia globosa, a small-sample regime where bias is most severe, under matched budgets, a shared selection operator and an identical fitness function, with the cardinality prior controlled separately. Under conventional flat cross-validation the algorithms separate decisively (Friedman chi-square = 52.75, p = 2.1e-11) and hypervolume ranks NSGA-II first. Under nested cross-validation the same comparison is not significant (chi-square = 5.85, p = 0.119) and the rank order changes. Three controlled experiments identify the mechanism. A budget sweep shows optimistic bias rising monotonically with search effort (21.9% to 40.7% from 270 to 3030 evaluations) while honest error does not improve. A label-permutation null shows the strongest searcher recovering apparent signal from destroyed labels (RMSECV 7.96 against a no-skill value of 8.69), whereas weaker searchers do not. A landscape analysis finds 386 solutions within 10% of the optimum with median pairwise Jaccard 0.056, including disjoint pairs of equivalent fitness. This degeneracy is an engineering asset rather than a defect, since many channel sets reach the same accuracy and the filter set of a low-cost instrument can be chosen on cost or manufacturability grounds. Hypervolume computed on estimated objectives therefore separates algorithms along a dimension uncorrelated with unbiased performance, search intensity converts into selection bias, and the low selection stability routinely reported reflects structural non-identifiability rather than algorithmic deficiency. A minimum reporting protocol is proposed. |
