In this paper, a novel subspace learning method, semi-supervised marginal discriminant analysis (SMDA), is proposed for classification. SMDA aims at maintaining the intrinsic neighborhood relations between the data points from the same class, while maximizing the margin between the neighboring data points with different class labels. Different from traditional dimensionality reduction algorithms like linear discriminant analysis (LDA) and maximum margin criterion (MMC) which seeks only the global Euclidean structure, SMDA takes local structure of the data into account. Moreover, it is designed for semisupervised learning which incorporates both labeled and unlabeled data points and avoids suffering the small sample size (SSS) problem. QR decomposition is then employed to find the optimal transformation which makes the algorithm scalable and more efficient. Experiments on face recognition are presented to show the effectiveness of the method.