Abstract: Subspace learning is an effective dimensionality reduction method. However, the resulting basis vectors are significantly biased due to the presence of outlier points. Consequently, the transformed data in the subspace cannot faithfully describe the intrinsic distribution of the original data. To tackle this problem, a modified subspace learning algorithm is proposed. In the algorithm it is not necessary to detect outliers. Moreover, the algorithm is reduced to an eignenvalue problem which has a globally optimal solution. Experiments on synthetic data demonstrate the effectiveness of the proposed algorithm.