The increasing adoption of point clouds in the digital documentation of Cultural Heritage (CH) has made three-dimensional semantic segmentation a key step for data interpretation and analysis. In this context, Deep Learning (DL) approaches have demonstrated high performance, albeit at the cost of substantial computational requirements and the need for large annotated datasets. Within this framework, the present study investigates the potential of leveraging a traditional supervised Machine Learning (ML) approach - Random Forest (RF) - through targeted optimization of training and validation procedures. To this end, the RF_CHC (Random Forest for Cultural Heritage Classification) model is proposed. Aimed at improving accuracy and, in particular, generalization capability in the semantic classification of architectural CH point clouds, RF_CHC integrates statistical hyperparameter calibration through the adoption of cross-validation procedures. The performance of RF_CHC was evaluated and compared with literature models (RF4PCC and optimized RF4PCC), demonstrating improved classification consistency and greater robustness across heterogeneous datasets, while highlighting the potential of an optimized ML-based approach as a competitive or complementary solution to currently prevalent DL models in the CH domain.