<p>The accuracy of radial basis function (RBF) interpolation is highly sensitive to the choice of the shape parameter. Existing methods for selecting this parameter are mostly based on empirical rules or optimization algorithms, which often suffer from poor generalization ability and high computational cost. Based on the error theory of RBF interpolation, this study investigates the intrinsic relationship between the geometric characteristics of the target function and the shape parameter. A set of geometric features closely associated with the shape parameter is extracted from extensive experimental data. On this basis, a Stacking regression model is constructed to develop a predictive approach for determining the shape parameter in inverse multiquadric RBF interpolation. Numerical experiments comparing the proposed model with simulated annealing, AdaBoost, and random forest demonstrate that the proposed method achieves superior interpolation accuracy across all tested functions, offering a new perspective on shape parameter selection in RBF interpolation.</p>