CURE-SMOTE algorithm and hybrid algorithm for feature selection and parameter optimization based on random forests.

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Random Forests (RF) algorithm offers broad applicability and overfitting resistance.
• Existing RF methods face challenges with imbalanced data, feature selection, and parameter optimization.
• This research aims to enhance RF performance by addressing these limitations.

Purpose of the Study:

• To develop an improved method for imbalanced data classification using Random Forests.
• To introduce novel techniques for feature selection and parameter optimization within the RF framework.
• To evaluate the effectiveness of proposed algorithms against existing methods.

Main Methods:

• Proposed the CURE-SMOTE algorithm, integrating Clustering Using Representatives (CURE) with Synthetic Minority Oversampling Technique (SMOTE) for imbalanced data.
• Developed a hybrid Random Forests (RF) algorithm utilizing minimum out-of-bag (OOB) error for feature selection and parameter optimization.
• Investigated hybrid RF algorithms incorporating genetic, particle swarm, and fish swarm optimization.

Main Results:

• CURE-SMOTE demonstrated superior performance in imbalanced data classification compared to standard SMOTE variants and random sampling.
• Hybrid RF algorithms achieved minimum OOB error, indicating enhanced generalization ability on binary and high-dimensional datasets.
• The proposed hybrid algorithms outperformed the original RF algorithm in terms of F-value, G-mean, AUC, and OOB scores.

Conclusions:

• CURE-SMOTE generates training sets with reduced noise, leading to improved classification accuracy.
• The hybrid RF algorithm offers a novel and effective approach for feature selection and parameter optimization.
• The developed methods provide significant advancements in Random Forests' application to complex datasets.