In this study, we propose an interaction identification method for classification tasks involving ultra-high dimensional discrete binary features.  We introduce a parameter to characterize feature interactions, then derive an estimator for this parameter and provide a theoretical proof of its asymptotic normality.  Based on this asymptotic distribution, hypothesis testing is conducted to determine whether a given feature pair exhibits significant interaction. To address the multiple-testing challenge arising from the vast number of hypotheses, we employ the e-BH procedure to screen and select significant interacting feature pairs. Theoretical guarantees are established to ensure the screening procedure’s consistency and control of false discovery rates. Finally, by generalizing the classical naive Bayes classifier, we propose an e-BH naive Bayes classifier that integrates identified feature interactions. Numerical simulations and real-world data experiments demonstrate the superior performance of our proposed methodology in capturing complex feature dependencies and enhancing classification accuracy.