The widespread use of online learning platforms has highlighted the challenge of managing student engagement and cognitive load. Poor interface design, complex layouts, dense contents, and redundant elements can increase cognitive strain, hindering learning effectiveness. To address this, adaptive interfaces based on user models are essential for optimizing cognitive load management in online learning environments. This study employs a mixed-method approach, focusing on user interaction behaviors as cognitive load indicators. Quantitative data were collected from 33 computer science students at a public university in Indonesia using the Hotjar service on an LMS. Participants engaged in a 150-min asynchronous learning session involving reading learning materials, watching instructional videos, participating in discussions, and taking quizzes. Learners’ cognitive load was assessed using the NASA-TLX questionnaire. Findings reveal that key interaction metrics—access duration, text input, clicks, and U-turns—correlate with cognitive load levels. Over 80% of respondents reported high cognitive load, with the most engaged learners experiencing the highest strain. K-means clustering identified four learner profiles, showing that intense interaction does not always lead to efficient learning. Feature importance analysis confirmed text input and access duration as the strongest predictors of cognitive load. This study highlights the need for adaptive learning systems that dynamically adjust user interface, content, and complexity based on real-time behavioral indicators. Future research should explore real-time adaptation, physiological data integration, and personalized learning pathways to enhance cognitive efficiency in online education.