Decision variable grouping methods in Multi-Objective Evolutionary Algorithms (MOEAs) play a crucial role in optimization tasks, yet existing approaches face inherent trade-offs between efficiency and accuracy. While some grouping strategies demonstrate computational efficiency, they often neglect the interdependencies among decision variables. Conversely, other methods capable of capturing variable correlations tend to incur substantial computational overhead and exhibit sensitivity to variable perturbation frequencies, potentially leading to inconsistent results during repeated applications. To address these limitations, this study proposes a novel gradient-based grouping methodology that effectively balances computational efficiency with dependency analysis. The core innovation lies in constructing a mathematical model through partial derivative computation of objective functions with respect to decision variables, enabling precise characterization of nonlinear interactions while maintaining computational efficiency.