Standard machine learning methods typically rely on intrinsic data features - essentially the absolute coordinates of samples in feature space - while overlooking the discriminative context provided by relative spatial relationships and data uncertainty. To address these limitations, we propose a novel feature augmentation framework termed Three-Way Feature Encoding with Shadowed Prototypes (TFSP), which serves as a concrete implementation of Shadowed Sets and Three-way Decision theory in representation learning. Beyond merely utilizing intrinsic data features, TFSP explicitly characterizes the relative spatial relationships between samples and class-specific prototypes. By encoding the relationships into ternary codes, TFSP provides a complementary context that enriches the feature representation and enhances the discriminative power. Furthermore, by incorporating the uncertainty modeling mechanism of shadowed sets and three-way decisions, the fused representation creates a safety buffer for ambiguous samples. This strategy effectively mitigates the risk of misclassification in boundary regions, thereby substantially improving model robustness. Crucially, the proposed encoding is independent of specific learning algorithms and can be seamlessly integrated into both statistical and deep learning methods to consistently boost classification performances. Extensive experiments demonstrate that TFSP significantly enhances classification accuracy and robustness, particularly in scenarios characterized by high data uncertainty.