<p>Machine learning algorithms working on fuzzy data generally rely on arithmetic operations derived from the classical expansion principle. However, repeated calculations on fuzzy numbers lead to excessive uncertainty accumulation, known as the "exploding fuzziness problem," which severely degrades model performance and numerical stability. In this study, we propose a class of constrained fuzzy arithmetic operations designed to control the accumulation of fuzzy in large-scale learning processes. The proposed framework is formally developed for Polynomial Triangular Fuzzy Numbers (PTFNs), and various theoretical properties, including closure and distance consistency, have been determined. Furthermore, the proposed operations have been integrated into a Fuzzy Gradient Boosting Regression model to evaluate their practical effectiveness. Experimental results on reference datasets demonstrate that the proposed approach successfully prevents fuzzy explosion and provides stable and accurate predictions compared to classical fuzzy arithmetic methods. These findings demonstrate that constrained fuzzy arithmetic offers a reliable computational framework for machine learning applications involving fuzzy-valued data.</p>