The recent rise of online learning and practice platforms has not only given the promise to close the enormous current gap in educational resources around the world, but also put forward the need for a better understanding of their effectiveness and contributing factors. Luckily, such investigation is made possible by platforms that carefully collect data on user behavior and content features. Published by an online IT practice platform called Educoder, MOOPer is a large-scale dataset consisting of user interaction data with the learning materials and structured side information of all these materials. Applying linear analysis and deep knowledge tracing models to MOOPer, this study examines the relationship between learners’ behavior and learning outcomes on the platform. More specifically, the amount of attempts required by a learner to solve an exercise is predicted by content category, difficulty, learning history, and various characteristics during their interaction, including the number of retry attempts, hint usage, and time spent. An encoder-decoder transformer model achieves the best predictive performance, while the linear regression analysis reveals the effect of active help-seeking behaviors on positive learning. The findings will help identify patterns associated with effective learning from learners’ behavior data. This information provides opportunities to adjust the course design to the needs of a variety of learners, ultimately enhancing their learning experiences and outcomes.