Aimed at the problems that timeliness is poor in geometric flight conflict detection methods and detection samples are imbalanced in machine learning-based detection methods, this paper proposes a flight conflict detection method of integrating spatial geometric modeling and machine learning. Firstly, in combination of synthetic multi-dimensional features such as aircraft position and velocity, the geometric method is used to determine conflicts based on the velocity obstacle method and the three-dimensional cylindrical protection zone. In order to address the poor timeliness of geometric judgment, a machine learning method is introduced. Conflict samples being short and training samples being imbalanced, the lightweight Gradient Boosting Machine (LightGBM) algorithm with a class weight adjustment mechanism is selected. Finally, the proposed method is verified by using actual secondary radar data collected in the Xi’an area. The experimental results show that the operating speed by the proposed method is 3.91 times faster than that by the geometric method. Compared with typical algorithms such as Random Forest (RF) and K-Nearest Neighbors (KNN), the proposed method in conflict judgment is an accuracy increase of 19% and 91% respectively.