Dynamic cost-sensitive random forest for imbalanced data

doi:10.1007/s10586-026-06245-z

Dynamic cost-sensitive random forest for imbalanced data

2026-07-13

0

PRE

AI

M

Mohammad-Awis Al-Lababede *

N

Nailah Al-Madi

DOI：10.1007/s10586-026-06245-z

Original

Original request for help

Save

Abstract

En 中文

Random Forest (RF) has demonstrated high effectiveness and accuracy in many domains, especially when the data used is imbalanced. One of the approaches to deal with imbalanced data is cost-sensitive learning, which focuses on the minority class and assigns it a higher cost compared to the majority class. Integrating cost-sensitive learning into Random Forest can enhance its performance for imbalanced data. However, determining appropriate cost values remains a challenging issue. This paper addresses this challenge through the use of dynamic costs. This paper proposes a novel cost-sensitive model called Dynamic Cost-Sensitive Random Forest (DCRF), based on the idea of incorporating dynamic costs into the RF model. The DCRF incorporates both traditional decision trees and cost-sensitive decision trees within the same ensemble using different proportions. For example, DCRF (0%) represents a model composed entirely of traditional decision trees, whereas DCRF (100%) represents a model composed entirely of cost-sensitive decision trees. Moreover, to achieve dynamic cost sensitivity, the DCRF generates N cost matrices, where N represents the number of decision trees; each cost matrix emphasizes the data from a different cost perspective. The proposed DCRF was evaluated on three imbalanced datasets using different cost scenarios. These datasets are: the Heart Disease Health Indicators dataset with an imbalance ratio of approximately 9.62:1, the Credit Card Fraud Detection dataset with an imbalance ratio of approximately 578.88:1, and the IT Customer Churn dataset with an imbalance ratio of approximately 2.77:1. Based on multiple evaluation metrics, including recall, precision, F1-score, accuracy, and Area Under the Curve (AUC), the model was compared against traditional RF, XGBoost, cost-sensitive decision trees (CSDT), and cost-sensitive random forests (CRF). The results show that DCRF achieved the best performance in terms of F1-score, indicating a better balance between precision and recall compared to RF, XGBoost, CSDT, and CRF. Additionally, DCRF achieved strong performance in terms of AUC, exceeding the results of XGBoost, CSDT, and many versions of CRF, while achieving performance comparable to that of RF.

Keywords:

Cost-sensitive

Random forest

Imbalanced dataset

Machine learning

Dynamic cost-sensitive random forest for imbalanced data

Abstract

Journal

Organization

Cited Papers