A new dual-scale nearest neighbor statistical feature construction algorithm for imbalanced data oriented to Gaussian...

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Gaussian Naive Bayes (GNB) classifier performance degrades on imbalanced datasets.
• Sparse minority class features and severe class overlap contribute to GNB performance issues.
• Existing methods like sampling and feature enhancement often mismatch GNB's core assumptions.

Purpose of the Study:

• Propose a novel feature construction algorithm, dynamic dual-scale nearest neighbor statistical ratio (NNDSR), to address GNB performance degradation on imbalanced data.
• Enhance the discriminability and Gaussian distribution adaptability of features.
• Improve the recognition accuracy of minority classes and overall classification performance.

Main Methods:

• Developed a dynamic dual-scale nearest neighbor mechanism to extract local aggregation and inter-class boundary information.
• Generated new features using cross-class and dual-scale statistical ratio operations.
• Validated the NNDSR algorithm on 22 UCI datasets with varying characteristics.

Main Results:

• NNDSR significantly outperformed the original data and 16 mainstream algorithms across key metrics (AUC, G-mean, F-measure).
• Demonstrated notable improvement in minority class recognition accuracy.
• Confirmed efficiency and stability through scalability tests on large datasets.

Conclusions:

• NNDSR is a robust feature construction algorithm for GNB to effectively handle imbalanced data.
• The proposed method avoids information distortion issues of traditional sampling techniques.
• NNDSR offers strong practical application value for imbalanced classification tasks.