Industrial Automatic Control Systems and Controllers Annotation << Back Geometry-Aware Construction of Reduced Training Sets for Tabular Classification Shayakberov E.V. The paper studies training set construction for classification on tabular data. The proposed approach partitions the centered numerical feature space into sign regions and selects rows within each region-class key. For correct scenario comparison, all scenarios within the same dataset are evaluated on identical external test parts, while only the training part is changed. Balanced accuracy is used as the main metric. Experiments were carried out on 21 public datasets. The best compact geometry-aware scenario improves balanced accuracy on 15 out of 21 datasets, keeps it at the baseline level on one dataset, and underperforms on five datasets. However, the effect is not universal: on some datasets geometry-aware selection outperforms random controls of the same size, while on others such an advantage is absent. The results also show that reducing the number of training rows affects epoch time much more often than full training time, namely on 19 out of 21 versus 6 out of 21 datasets. Keywords: tabular data, classification, training set reduction, sign regions, representative selection, balanced accuracy. DOI: 10.25791/asu.6.2026.1665 Pp. 19-28.