Business activity classification plays a crucial role in economic analysis, yet many classification processes remain manual and labor-intensive. Most previous studies rely on flat classification methods that do not leverage hierarchical label relationships, particularly in large-scale business taxonomies. Existing hierarchical approaches, such as Local Classifier per Level (LCL), improve scalability but often produce inconsistent predictions across hierarchical levels during inference. To address this limitation, we propose a hybrid inference strategy that reconciles top-down and bottom-up strategies through level-aware probability aggregation without requiring architectural modifications or additional training. Label Smoothing Regularization (LSR) is further incorporated to improve probability calibration during hierarchical probability aggregation. The proposed method is evaluated on the Indonesian Standard Business Classification (KBLI) taxonomy using a synthetic dataset with complete five-level hierarchical coverage. Experimental results show that the exponential-weighted hybrid strategy achieves the highest full path accuracy (FPA) of 0.9269 using IndoBERT with LSR, outperforming independent, top-down, bottom-up, and multi-task hierarchical classification approaches. Evaluation on real-world data further reveals substantial generalization challenges caused by semantic underspecification and distributional shift in short business descriptions. Overall, the findings demonstrate that inference-level optimization provides an effective and computationally efficient strategy for improving prediction consistency in automated business activity classification.
Noviyanti et al. (Mon,) studied this question.