Abstract: This study proposes a Bayesian optimization-based Kolmogorov-Arnold network (BO-KAN) for rockburst intensity prediction to improve predictive accuracy while preserving feature interpretability in deep underground engineering. The KAN architecture is introduced, with learnable activation functions on the edges parameterized by B-spline basis functions, and its symbolic representation capability is exploited to build a prediction model that is both accurate and interpretable. Bayesian optimization is used to tune key hyperparameters, including the L1 penalty, entropy regularization, and learning rate, thereby enhancing the model’s generalization ability. Experimental validation on 65 engineering cases showed that the BO-KAN model achieved 90% accuracy on the test set and outperformed several conventional machine-learning methods. In addition, symbolic extraction identified statistically meaningful class-discriminant functions from the input features and revealed six intrinsic mechanisms—suppression, dominance, eruption, marginality, extreme events, and oscillation—thereby linking data-driven prediction with feature interpretability. The proposed method offers a new approach to rockburst risk prevention and control in deep underground engineering.