Educational process mining is now a promising method to provide decision-support information for the teaching-learning process via finding useful educational guidance from the event logs recorded in the learning management system. Existing studies mainly focus on mining students' problem-solving skills or behavior patterns and intervening in students' learning processes according to this information in the late course. However, educators often expect to improve the learning outcome in a proactive manner through dynamically designing instructional strategies prior to a course that are more appropriate to students' average ability. Therefore, in this article, we propose a two-stage problem-solving ability modeling approach to obtain students' ability in different learning stages, including the pre-problem-solving ability model and the post-problem-solving ability model. The models are trained with Gradient Boosting Decision Tree (GBDT) on the historical event logs of