Taking Olympic spirit-ideological and political education as an example, this paper proposes a three-stage practical teaching model based on machine learning algorithms to optimize the interaction strategy from the background of the actual demand of practical teaching online. To capture and predict the learning state of students, an online education prediction model has been established. The LightGBM model and labeling approximation step are introduced in two structures, namely, boosting trees and convolutional neural networks, respectively. The application of the model is analyzed in terms of learning attitude and satisfaction, learning performance, and interactive behavior, and experimental conclusions are obtained. After applying the three-stage interactive Civics teaching model, the difference between the pre- and post-test Civics scores of the experimental group is significant (Sig.=0.001). The average scores of each item of students’ attitude and satisfaction are more than 3.98. The students’ participation in this paper’s model is as high as 66.27%, in which the highest frequency of interaction is student-student interaction (40.7%), followed by teacher-student interaction (25.86%). Therefore, the model of this paper can improve the initiative of learning and enhance the frequency and depth of interaction.