<![CDATA[The classification of bandgap types in perovskite ABO₃ materials presents a significant challenge due to overlapping feature distributions and class imbalance between direct and indirect bandgap classes. This study investigates the impact of data cleaning using the Edited Nearest Neighbour (ENN) method on the performance of multiple machine learning classifiers, including Multilayer Perceptron (MLP), Gradient Boosting, CatBoost, and Extra Trees. Model evaluation focuses on accuracy and F1 macro to capture both overall performance and class wise balance. Under baseline conditions, the best performing model achieved a test accuracy of 0.8934 with an F1 macro of 0.7783, indicating strong predictive capability but limited sensitivity to the minority class. After applying ENN exclusively to the training data, the class distribution became more balanced, resulting in improved recall for the minority class across several models. However, this improvement was accompanied by a reduction in overall accuracy. The highest F1 macro under ENN was 0.7632, achieved by the Gradient Boosting model, demonstrating enhanced class balance despite a lower accuracy of 0.8646. Cross validation results confirm that ENN does not significantly increase performance stability but effectively refines decision boundaries by removing ambiguous majority class samples. These findings highlight the trade off between accuracy and class fairness and emphasize that model selection should be guided by the evaluation objective. For imbalanced bandgap classification tasks, F1 macro oriented evaluation combined with ensemble based models provides a more reliable and equitable predictive framework.]]>
Copyrights © 2026
