An enhancement of kavya et al.’s random forest algorithm applied for flood predictions in Sampaloc, Manila

ABSTRACT: Kavya et al.’s Random Forest Algorithm was used for allergy disgnosis using datasets sourced from allergy testing center in South India, effectively addressing cases with comorbidities. However, the algorithm faces a challenge due to the presence of noisy data in datasets, which may result in poor performance. This study explores the limitations of Kavya’s Random Forest Algorithm, particularly overfitting and bias toward majority classes and proposes the integration of Recursive Feature Elimination (RFE), Nearmiss Undersampling Method, and Bayesian Optimization to enhance its predictive reliability. The enhanced algorithm incorporates NearMiss as an undersampling technique to address class imbalance, RFE to eliminate redundant features, and reduce noise, and Bayesian Optimization for efficient hyperparameter tuning. Simulations using five datasets demonstrated substantial performance gains. On average, the enhanced algorithm improved accuracy and recall by approximately 57%, and F1-score by 73% compared to the existing algorithm. Precision increased by an average of 38%, while ROC AUC improved by an average of 30%, indicating better class separation. Additionally, the overfitting gap was significantly reduced, from an average of 60% to just 1.3%, demonstrating improved generalization and stability across datasets. These results indicate that integrating RFE, NearMiss, and Bayesian Optimization effectively mitigates overfitting and improves the model’s robustness, making it a more reliable tool for flood prediction. The enhanced algorithm provides a more dependable solution, with its enhanced performance making it more appropriate for flood prediction, aiding in disaster management and preparedness.

Filipiniana