Faced with the diverse preferences of students, traditional education systems often struggle to deliver personalized learning experiences. Intelligent Tutoring Systems (ITSs) have emerged as a vital tool to enhance educational engagement. While existing AI-powered educational approaches have made some progress, they still fall short of meeting the diverse and constantly evolving needs of students—especially those whose preferences are not yet clearly defined. This paper proposes a Personalized Learning Style Classification and Adaptive Teaching (SCAT) model. SCAT integrates advanced architectures including Swin-Transformer, ResNet, and EfficientNet to enable dynamic classification of learning styles and personalized instructional strategies. To further enhance model performance, this paper introduces novel data processing techniques: adaptive range scaling, feature fusion and weight adjustment, and uncertainty-driven transformation. Comprehensive evaluation across multiple metrics demonstrates that SCAT effectively handles complex and imbalanced datasets, as well as temporal dependencies in learning behaviors. It accurately identifies nuanced learning style preferences across diverse student groups. In future, SCAT can serve as a powerful component of ITS, facilitating real-time instructional adaptations based on students’ engagement patterns.