To reduce the impact of noise and outliers on ensemble pruning and robustly select a more sparse subset of base classifiers to improve the performance of facial expression recognition, a robust sparse and low-redundancy ensemble pruning method with dependency scores is proposed in this paper for facial expression recognition. First, the method treats the prediction results of sample instances as base classifier features, and evaluates the dependency and priority between pairs of base classifiers using mutual information and entropy, respectively. Second, the priority dependency is added to the regression-based objective equation to prune redundant base classifier, which uses the l_2,1-norm to increase the robustness of the classifier subset and thus improve the generalization performance of the algorithm. Then, an inner product regularization term is introduced into the target equation to select sparse and low-redundant base classifiers by computing the sum of the absolute values of the inner products of the classifier feature coefficient vectors. Finally, the majority voting method is used to integrate the selected subset of base classifiers to obtain the final recognition results. The recognition accuracy of the proposed method on four public facial expression datasets, FER2013, JAFFE, CK+, and KDEF, is 3.29%, 10.39%, 1.76%, and 4.89% higher than those obtained by integrating all base classifiers, respectively, indicating that the method can select a subset of classifiers with better recognition results and lower redundancy, and improve the ensemble pruning generalization ability.