Abstract
Abstract
A common step in unsupervised domain adaptation is to remove an estimated domain sub-space from a learned representation, assuming that it is a nuisance rather than a class-relevant signal. However, this is rarely tested directly. Using deployment-shift screening of germinated oil palm seeds, we propose a label-free diagnostic, leveraging the angle between the domain subspace and the source class-discriminative direction and the cross-representation stability of that direction, to test statistically whether it tracks the removal effect. In a matched, causal design consisting of 1,105 configurations (over three feature spaces, three targets of increasing severity, three seeds, and each paired with a raw counterpart), the matched-pairs statistics, correlation and interaction tests confirm the association: removal is neutral to mildly harmful under mild shift, small but consistently positive under moderate shift, and largest under severe shift (0.10–0.15 balanced accuracy), always in the direction the diagnostic indicates. The association is coarse: it indicates which representation is worth trying, not how large the gain will be, and the residual angle–effect correlation is null once representation and target are controlled for. A retrospective oracle confirms that strong operators exist for every target, but source-validation selection grows markedly less reliable under severe shift, precisely where the potential gain is largest.