ROBUST ESTIMATION METHODS FOR k-NEAREST NEIGHBOURS ENSEMBLE REGRESSION MODEL

Abstract

The k-nearest neighbor (k-NN) regression method is widely used due to its simplicity and flexibility; however, its reliance on mean-based neighborhood aggregation makes it highly sensitive to outliers, noise, and skewed data distributions. To address these limitations, this study proposes four robust extensions of the k-NN regression framework: Median k-NN ensemble (MKNNE), Winsorized k-NN ensemble (WKNNE), Trimean k-NN ensemble (TriKNNE), and Trimmed Mean k-NN ensemble (TKNNE). By replacing the conventional sample mean with robust measures of central tendency, the proposed methods enhance robustness without sacrificing computational efficiency. Extensive experiments conducted on ten benchmark datasets with diverse statistical properties demonstrate that the proposed models consistently outperform standard k-NN, random k-NN (RKNN), and optimal k-NN ensemble (OKNNE) methods across multiple evaluation metrics, including R², MSE, MAE, and MAPE. The results show that robust neighborhood aggregation is an effective strategy for improving k-NN regression, especially in real-world scenarios involving noisy and heterogeneous data. This work provides a robust and extensible framework for neighborhood-based learning.