Designing an AI-Driven Personalization Framework for E-Learning: A Design Science Research Study

The increasing popularity of e-learning and Learning Management Systems (LMS) have led to expectations of highly customizable learning environments tailored not only to the knowledge and behavior of the learner but also to his/her characteristics and level of engagement. When talking about learner profiling, Learning Analytics, knowledge tracing, and personalized recommendation, AI is very frequently used. On many occasions, these AI capabilities are explored as individual features only. Through a Design Science Research (DSR) method, this paper bridges the gap between these different AI personalization techniques by proposing an AI-Driven Personalization System (ADPF) for e-learning. To support the system design, the research team identified 383 publications on five academic databases, of which 28 have been reviewed for this paper as they directly pertain to setup design. This study extracted from the literature 7 design requirements for the structure covering learner profiling, behavioral analytics, knowledge estimation, personalized recommendation, dynamic adaptation, continuous feedback, and LMS compatibility. These seven requirements have been grouped into five different setup components, which are: Learner Profiling, Learning Analytics, Knowledge Tracing, Personalized Recommendation, and Adaptive Learning Interface. The learning analytics part of a model collects learner interaction with learning content as data and processes these interactions. This model offers a learning path recommendation and gives dynamic content adaptation for individualized learning to occur. At the same time, learner profile is continually updated using learner activity data and the model is adjusted as a result. Also, the model is updated continuously and the learner is fed back about the updates. The main benefit of the model was that it was an architectural solution that integrated personalization features of different AI techniques in one system and offered the basis of the first physical realization and an empirical investigation. Scenario-based analysis further showed the setup's ability to deal with varied learning conditions of learners such as low competencies, mastering, and dropping engagement.