Abstract: Study session dropout prediction allows for educational systems to identify when a student would stop a study session which gives vital information to prolong learning activity. Student session dropout can depend on many factors that are involved with the engagement when using the system. The student’s knowledge level and their track records within the system are closely related to the student’s willingness to continue with their studies. Knowledge tracing as a task models the user’s knowledge level given study history. The information from knowledge tracing can have significant impact on predicting the student’s willingness to continue, which is why it is natural to train two tasks jointly for better generalization in dropout prediction task. While extensive research has been conducted individually on dropout prediction and knowledge tracing, the effect of jointly modeling two tasks has not been thoroughly investigated. Hence, we show that multi-task training of the study session dropout prediction model along with knowledge tracing boosts the performance of study session dropout prediction, especially on more challenging tasks and datasets. Specifically, with Transformer-based models, multi-task training significantly improves Area Under Receiving Operator Curve (AUROC) by 3.62% in further N-step dropout prediction task, which is a study session dropout prediction task under a more practical setting. Moreover, under label-scarce and class-imbalance settings, our method shows improvements of AUROC up to 12.41% and 11.22%, respectively. Our results imply that knowledge tracing is closely related to study session dropout prediction and can transfer positive knowledge in multi-task training, which provides a new way to better predict dropouts especially in difficult settings.