DualSelect: Active learning by structural pseudo-label instability under complementary support perturbations

In pseudo-label-driven active learning, uncertainty conflates two situations: a sample the model merely hesitates on, and one whose pseudo-label is sensitive to the surrounding unlabeled structure. We formalize the latter as structural pseudo-label instability (SPI): the Jensen–Shannon disagreement between classifiers trained on complementary cluster-balanced halves of the pseudo-labeled pool. Query-by-Committee varies the hypothesis at fixed data; DualSelect fixes the learner and perturbs the local support, needing no independent feature views. SPI is selective rather than universal: its average ablation effect is small, with gains concentrated under residual class ambiguity. DualSelect therefore combines SPI with uncertainty and geometry through data-conditioned weights of fixed functional form, a representation-aware gate, and candidate diversification. The ablation attributes performance to this coordination rather than to SPI alone: removing the gate costs 2.0–2.7 pp of area under the learning curve on every family. Fourteen competitors are compared across four regimes: synthetic data, eleven UCI datasets at two query granularities, and latent image representations, with a second backbone on the synthetic campaign. The claim is robustness, not uniform dominance. On the tabular benchmark DualSelect trails the best mean by 0.09 and 0.07 pp, yet wins more dataset–budget cells than any competitor and is significantly outperformed by none. On the latent benchmark it is significantly ahead of all 14 competitors. Across the four regimes it has the lowest mean and worst-regime empirical regret (0.74 and 2.08pp), and is the only method placing among the top three in every regime.