A Comparative Study of Structural Improvement and Hyperparameter Optimization for Enhancing Naïve Bayes Performance

Abstrak

The Naïve Bayes (NB) algorithm is widely used due to its simplicity and computational efficiency; however, its performance often degrades on real-world data because the assumption of feature independence is frequently violated. This study evaluates two strategies for improving the performance of Naïve Bayes, namely hyperparameter optimization and feature dependency structure enhancement through Tree Augmented Naïve Bayes (TAN). The Breast Cancer Wisconsin (Diagnostic) dataset was selected because it is a representative and widely used public dataset for evaluating medical classification algorithms, thereby facilitating method validation and comparison with previous studies. Experiments were conducted on 30 numerical features with an imbalanced class distribution (62.74% Benign and 37.26% Malignant). The baseline Naïve Bayes model achieved an accuracy of 0.9386. Applying TAN as a standalone approach improved performance to an accuracy of 0.9474 and an F1-score of 0.9286. In contrast, hyperparameter optimization using Genetic Algorithm (GA), Grid Search (GS), and Ant Colony Optimization (ACO) without TAN did not yield meaningful performance improvements; in fact, GA resulted in a decrease in accuracy compared to the baseline model. The combinations of Naïve Bayes + TAN + GS and Naïve Bayes + TAN + ACO achieved the best performance, with identical metrics: accuracy of 0.9561, precision of 0.9744, recall of 0.9048, F1-score of 0.9383, and AUC of 0.9960. These results indicate that improving the feature dependency structure through TAN has a more fundamental impact on enhancing Naïve Bayes performance than hyperparameter optimization alone.

Identifier

• OpenAlex: https://openalex.org/W7160916898
• DOI: https://doi.org/10.30871/jaic.v10i2.12278