An improved SMOTE algorithm for enhanced imbalanced data classification by expanding sample generation space.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Class imbalance in datasets significantly degrades classification model performance.
• Existing over-sampling techniques like SMOTE often fail to preserve local data density and distribution.
• Improved methods are needed to synthesize samples that better reflect original data characteristics for robust classification.

Purpose of the Study:

• To introduce an enhanced SMOTE algorithm (ISMOTE) that incorporates local spatial information for synthetic sample generation.
• To address the limitations of traditional SMOTE in handling local data distribution and density distortions.
• To improve the robustness and performance of classification models on imbalanced datasets.

Main Methods:

• Proposing ISMOTE, which modifies spatial constraints for generating synthetic samples.
• Generating a base sample between two original samples and using Euclidean distance to create new samples.
• Adaptively expanding the synthetic sample generation space to better preserve local data distribution.

Main Results:

• Comparative analysis with seven over-sampling algorithms on thirteen public datasets.
• ISMOTE demonstrated more realistic data distributions in 2D and 3D scatter plots.
• Significant improvements in classifier performance: F1-score (+13.07%), G-mean (+16.55%), and AUC (+7.94%).

Conclusions:

• ISMOTE effectively alleviates distortions in local data distribution and density.
• The algorithm shows parameter adaptability for multi-class imbalanced datasets.
• ISMOTE offers a superior approach to handling class imbalance for improved machine learning model performance.