A novel sparse Bayesian learning for correlated error (SBL-CE) algorithm is proposed to automatically search for an optimal model class with relevance features in regression problems of pattern recognition based on measured data and extracted features. The proposed SBL-CE algorithm is designed to overcome the disadvantage in the traditional optimal model searching approach for ground motion pattern recognition, which requires a huge or even intractable computational effort to examine a large number of different combinations of extracted features. The proposed SBL-CE algorithm introduces sophisticated hyperparameterization on the regression parameter vector in the ground motion prediction model, aiming to conduct a continuous optimal model search even when the number of extracted features is large. In addition, the prediction error independence assumption in the traditional learning approach is relaxed, so the derived optimization strategy can be applied to ground motion pattern recognition. The proposed SBL-CE algorithm is then used to analyze a database of strong ground motion records in the Tangshan region of China. It is shown that the model by the proposed SBL-CE algorithm is superior compared to the traditional models because it is capable of properly recognizing the pattern of ground motion in the target seismic region with high accuracy and robustness.