To address the problem of unsatisfactory recognition results in cross-database micro-expression recognition due to inconsistent distribution of features between training and testing samples, a domain generalization method for cross-database micro-expression recognition based on Feature Iteration Selection (FIS) is proposed. FIS iteratively discards the features that have a large impact on the classification results of the model in the training phase, forcing the network to activate the remaining features to participate in training, preventing the fully connected layer from predicting with the most predictive subset of features, and balancing the strength of the model in extracting information from different features, thus improving the generalization ability of the model . Experiments on three widely used micro-expression datasets show that the FIS method achieves an average accuracy of 54.54% and an average F1 value of 54.20%, outperforming the mainstream domain adaptive and domain generalization methods, validating the superiority of the proposed FIS method for cross-database micro-expression recognition tasks.