Personalized learning, i.e., recommending personalized remediation or enrichment activities to each learner based on their individual background, interests, learning goal, and learning progress, has the potential to revamp the outdated “one-size-fits-all” approach in education. The key to achieving scalable personalized learning is to leverage recent developments in machine learning (ML) and big data systems to collect and analyze learner and content data and automatically generate actions for personalization. Developing ML methods for personalized learning comes with significant challenges since (i) it is difficult to design interpretable and robust learner models due to the complex nature of human behavior, and (ii) the outcomes of education (e.g., employment) are of paramount importance to the learners, which places a premium requirement on the safety of personalization. In this talk, I will propose novel ML methods that improve the interpretability and safety of personalized learning. First, I will introduce a series of learner-response models that offer superior interpretability on the knowledge structure of assessment questions. Then, I will propose a linearization method that provides exact, nonasymptotic expressions of the error in learner and content parameter estimates, which is critical to the design of fail-safe personalization algorithms.