In face recognition (FR), a lot of algorithms just utilize one single type of facial features namely global feature or local feature, and cannot obtain better performance under the complicated variations of the facial images. To extract robust facial features, this paper proposes a novel Semi-Supervised Discriminant Analysis (SSDA) criterion via nonlinearly combining the global feature and local feature. To further enhance the discriminant power of SSDA features, the geometric distribution weight information of the training data is also incorporated into the proposed criterion. We use SSDA criterion to design an iterative algorithm which can determine the combination parameters and the optimal projection matrix automatically. Moreover, the combination parameters are guaranteed to fall into the interval [0, 1]. The proposed SSDA method is evaluated on the ORL, FERET and CMU PIE face databases. The experimental results demonstrate that our method achieves superior performance.