Car Acceptability Classification: A Comparative Study of Flat and Hierarchical Rule-Based AI Systems

The transparency of rule-based artificial intelligence makes it attractive for decision tasks, yet the relationship between model structure, predictive performance and human interpretability is not straightforward. This study compares two rule-based classifiers on the UCI Car Evaluation Dataset (1,728 categorical records): a flat sequential IF–THEN decision list and a hierarchical symbolic model that combines PRICE, COMFORT and TECH intermediate concepts. Models were evaluated using a stratified 346-record held-out test set, with macro-F1 designated as the primary metric, alongside accuracy, class-specific results and explanation structure. On the held-out set, hierarchical rules obtained macro-F1 of 0.7652 and accuracy of 0.8873, compared with 0.7323 and 0.8786 for flat rules. However, across five exploratory paired splits, flat rules achieved higher mean macro-F1 (0.7615 versus 0.7249), leading in four splits. The hierarchy improved recall for the good and vgood classes on the original holdout but did not establish uniformly better minority-class performance. Because the approaches also differ in learning strategy and information representation, their performance differences cannot be attributed solely to architecture. The study demonstrates sampling sensitivity and contrasts structural explanation formats without claiming measured improvements in human understanding.