A unified approach to robust classification in the presence of outliers

Outlier detection in real-world machine learning remains challenging due to noisy samples, borderline instances, and heterogeneous data distributions. Traditional methods typically rely on either hard removal or uniform sample weighting, which may discard informative yet uncertain observations and consequently degrade classification performance. In addition, most existing approaches treat sample influence, uncertainty, and local density as separate components rather than incorporating them into a unified learning framework. In this work, we introduce a unified influence–uncertainty–density learning framework that simultaneously models global sample influence, local boundary ambiguity, and neighborhood density through an adaptive reweighting mechanism. Specifically, global influence scores are estimated using influence functions to quantify sample importance, fuzzy clustering is employed to quantify feature-space uncertainty, and Local Outlier Factor is employed to capture local density characteristics for density-aware outlier estimation. These components are integrated into a unified learning framework that adaptively adjusts the contribution of each sample during classifier training without modifying the original feature representations. We evaluate the proposed framework on nine benchmark datasets and observe consistent performance improvements across multiple classifiers. For example, XGBoost achieves 98.99% accuracy on Wilt, 98.98% on Musk, 94.80% on Spambase, and 88.63% on Biodeg. Experimental results demonstrate that jointly modeling influence, uncertainty, and density enhances robustness against noisy observations while retaining informative borderline samples, leading to more stable and accurate classification performance across diverse datasets.