This article introduces MML-MAO (minimalist machine learning-Mexican axolotl optimization), a model that belongs to the recently developed MML paradigm. The underlying premise of this new paradigm is a version of Occam’s razor; in machine learning, effective, efficient, simple, and interpretable models are preferable. The conceptual and operational foundation of MML-MAO consists of selecting a single attribute from the total set of features present in the dataset. Then, through the application of the MAO metaheuristic, additional features are added to the initial simple attribute. For each training instance, the arithmetic mean of the small set of features selected by MAO is calculated; subsequently, by comparing this mean with a threshold, the class of the instance is determined. The experiments used 5-fold cross-validation as the validation method and the F1-score as the performance metric. The results of MML-MAO were compared with eight state-of-the-art machine learning algorithms across 22 datasets. According to the Friedman ranking, the MML-MAO model achieved the best overall performance; and according to the Bonferroni–Dunn test, it is statistically indistinguishable from SVM and statistically outperforms decision trees. The results confirm that compact, white-box models can outperform more sophisticated and complex models in less time and without the need for post-hoc interpretability tests. It is concluded that minimalist approaches deserve consideration by practitioners, especially in domains where transparency, interpretability, and scalability are essential.
Rodríguez et al. (Mon,) studied this question.