Data-driven computerized education platforms can drastically reduce costs of tutoring by statistically learning from the trajectories of the students. We present a reinforcement learning algorithm that is implemented in an online platform for recommending tutoring videos that are personalized to each student. For this purpose, multiple important challenges are addressed. First, the experiments for collecting data need to be diverse for exploring student responses, while at the same time are required to focus on the immediate weakness of each student. Moreover, the number of tutoring items is remarkably large, but each student provides an extremely small data because (s)he can engage in only a few items. Further challenges as well as employed methods that utilize student backgrounds for combining the data, yet recommending personalized tutoring, will be discussed.