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

Abstract

Abstract To address the performance degradation of Gaussian Naive Bayes (GNB) classifier on imbalanced datasets caused by sparse minority class features and severe class overlap, this paper proposes a new feature construction algorithm based on dynamic dual-scale nearest neighbor statistical ratio (NNDSR). The core of NNDSR is a dynamic dual-scale nearest neighbor mechanism, which is designed to accurately extract the local aggregation characteristics of samples and the inter-class boundary information. On this basis, new features are generated through cross-class and dual-scale statistical ratio operations. These features possess both strong discriminability and Gaussian distribution adaptability, which can significantly amplify class differences and effectively approximate the core assumptions of GNB. By optimizing the information expression of minority classes and enhancing class separability with these features, the algorithm avoids the information distortion problem of traditional sampling techniques and solves the mismatch between general feature enhancement algorithms and GNB’s core assumptions. Comparative experiments were conducted on 22 UCI datasets with varying scales, dimensions and imbalance ratios. Results show that NNDSR significantly outperforms the original data and 16 mainstream algorithms including sampling, feature enhancement and classifier-level optimization methods in core classification metrics such as AUC, G-mean and F-measure, with a notable improvement in the recognition accuracy of minority classes. Scalability tests further confirm its efficiency and stability on datasets with ten-thousand-level samples and within one hundred dimensions. This paper provides a robust new feature construction algorithm for GNB to handle imbalanced data, with strong practical application value.