This thesis focuses on personalized learning in education, a promising and effective means of learning where the instructions, educational materials, learning paths, analytics, and reports are tailored to each learner to best support their individual learning paths and improve learning outcomes. Current personalized learning relies heavily on expert instructors and is costly, has limited availability, and is unable to scale to meet the massive demand of learning today. This thesis takes a machine-learning approach to address the aforementioned issues by developing computational models that learn from educational big data to perform the activities central to personalized learning in education.

First, I will present a series of works for learning content customization, including methods and systems to generate, evaluate, represent, and analyze different types of learning content such as math word problems, factual quizzes, and scientific formulae. Second, I will present two frameworks for learning analytics, which enable the understanding and tracking of the progress of large numbers of learners effectively and efficiently. Finally, I will present methodologies for trustworthy machine learning, a necessity for deploying machine learning systems in real-world educational scenarios. These methodologies include theoretical tools for understanding recurrent neural networks, the powerhouse underlying modern knowledge tracing models, and controllable data generation, enabling machines to behave more precisely according to human instructions.