Network intrusion detection often suffers from the problem of imbalance, where a few classes of samples not only have poor accuracy and recall on their own, but also affect the overall accuracy and recall, for this reason, we use ADASYN (Adaptive Oversampling) and class balancing to balance the dataset. It also faces the problem of large data dimensions and dataset dispersion, for this reason, in this paper, we propose an intrusion detection method based on Recursive Elimination of Features (REF) with Logistic Regression Model with L1 Regularization and LightGBM (LGBM). Firstly, preprocessing is carried out by using solo thermal coding and label coding, after which the important features are retained by feature recursive elimination and the traffic features are dimensionality reduced. Then the optimal feature subset is passed into LightGBM for training prediction and the LightGBM algorithm is optimized using Bayesian algorithm. The experiments are validated using the public dataset NSL-KDD, and the final results show that the method proposed in this article has a significant improvement for the accuracy and F1-score, and also verifies the effectiveness of the method.