As a kind of common method for data preprocessing, feature selection can not only improve the classification performance, but also increase the interpretability of the classification results. In sparse-learning-based feature selection methods, some useful discriminative information is ignored, and it may affect the final classification performance. To address this problem, a new discriminative feature selection method called Discriminative Least Absolute Shrinkage and Selection Operator (D-LASSO) was proposed to choose the most discriminative features. In detail, firstly, the proposed D-LASSO method contained a L ₁-norm regularization item, which was used to produce sparse solution. Secondly, in order to induce the most discriminative features, a new discriminative regularization term was introduced to embed the geometric distribution information of samples with the same class label and samples with different class labels. Finally, the comparison experimental results obtained from a series of Benchmark datasets show that, the proposed D-LASSO method can not only improve the classification accuracy, but also be robust against parameters.