Enhancing Prediction Efficacy with High-Dimensional Input Via Structural Mixture Modeling of Local Linear Mappings