Weighted Model Fusion for Imbalanced Learning

Objectives: To develop a Weighted Multi-model Ensemble (WME) to improve binary and multiclass data predictions, particularly in handling imbalanced datasets. The model aims to achieve high performance metrics, such as precision and recall, while minimizing false positive rates. Additionally, the study seeks to explore better handling mechanisms to enhance prediction accuracy further. Methods: The methodology involves a two-phase approach for the Weighted Multi-Model Ensemble (WME). The first phase includes data preprocessing, segregating training and test data, and training models like Decision Tree, Random Forest, and Extra Tree Classifier to determine their accuracy and weights. The second phase applies the test data to these trained models, calculates probability levels, and makes final predictions based on the weighted probabilities. Results: The results indicate that the WME model achieved high True Positive Rate (TPR) levels, nearing 100%, and low False Positive Rate (FPR), demonstrating effective performance on the Brazilian bank dataset. Precision and recall values exceeded 90%, confirming the model's capability in identifying minority classes in moderately imbalanced data. Findings: The findings reveal that the Weighted Multi-Model Ensemble (WME) effectively handles data imbalance, achieving high precision and recall levels, particularly in the Brazilian bank dataset and KDD Cup 99 data. Novelty: The novelty of the study lies in the development of the Weighted Multi-Model Ensemble (WME), which integrates multiple classifiers with a unique weighting mechanism to enhance prediction accuracy on imbalanced datasets. This approach not only improves performance metrics like precision and recall but also addresses the challenges of minority class identification more effectively than existing models. Keywords: Data imbalance, Ensemble modeling, Weighted prediction, Probability based analysis, Classification