Akademik Veri Yönetim Sistemi
8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications, ICHORA 2026, Ankara, Türkiye, 21 - 23 Mayıs 2026, (Tam Metin Bildiri)
Feature Selection (FS) plays an important role in machine learning. Especially in time-series applications, large numbers of variables and overlapping information may negatively affect predictive performance. In this study, six FS techniques are examined under three main categories. The first is traditional approaches such as Mutual Information (MI) and Recursive Feature Elimination (RFE). The second is optimization-driven approaches that are represented by Particle Swarm Optimization (PSO) and Differential Evolution (DE). The third category includes explainability-oriented approaches, namely SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). Each method independently selects the best $k$ features, and the selections within each paradigm are aggregated through a set union to form three feature sets: FS-Set (combined sets generated by Traditional FS methods), Op-Set (combined sets generated by Optimization-based FS methods), and X-Set (combined sets generated by XAI-based FS methods). The resulting feature sets are evaluated using Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Multi-Layer Perceptron (MLP) classifiers in the IBTrACS tropical cyclone dataset for binary landfall prediction, an imbalanced time-series classification task. The results show that all three paradigms improve classification performance over the full-feature baseline, with the FS-Set achieving the highest F1-Score, the X-Set achieving the highest AUC-ROC, and the Op-Set providing the best trade-off between compactness and performance with the fewest features. The findings also reveal that different paradigms favor different classifier architectures, suggesting that the choice of FS strategy should be informed by the downstream model.