GQEO: Nearest neighbor graph-based generalized quadrilateral element oversampling for class-imbalance problem.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Class imbalance poses significant challenges to traditional classifier performance.
• Oversampling techniques, like SMOTE, augment minority instances but often rely on limited local information.
• Existing methods may overlook global neighborhood influences and are susceptible to noise.

Purpose of the Study:

• To introduce a novel oversampling technique, Generalized Quadrilateral Element Oversampling (GQEO), designed to overcome the limitations of local interpolation methods.
• To enhance the feature representation of minority instances by incorporating global neighborhood information.
• To develop a robust method that mitigates the impact of noise during data synthesis.

Main Methods:

• GQEO utilizes k-nearest neighbor (KNN) search to construct a global neighbor relationship graph.
• It identifies generalized quadrilateral elements within the graph, constrained by planar quadrilaterals.
• Minority instances are synthesized within these elements using finite element-inspired one-dimensional shape functions.

Main Results:

• Experimental results on 30 imbalanced datasets demonstrate GQEO's effectiveness in alleviating class imbalance issues.
• GQEO successfully prevents noise from influencing the instance synthesis process.
• The proposed method achieves competitive performance compared to state-of-the-art oversampling techniques, particularly those addressing minority noise.

Conclusions:

• GQEO offers a significant advancement in addressing the class imbalance problem by leveraging global neighborhood information.
• The technique provides a robust and effective approach to minority instance synthesis, outperforming existing methods in various scenarios.
• GQEO represents a promising direction for improving the performance of machine learning classifiers on imbalanced datasets.