Mobile Interaction Behavior Analysis for Optimizing Learning Engagement Prediction in Higher Education

Abstract

Mobile terminals serve as the primary carriers of self-directed mobile learning among university students. Interaction data generated from these terminals can authentically reflect learners’ dynamic engagement levels. Existing research, however, inadequately accommodates the non-uniform temporal characteristics of behavioral sequences, struggles to capture long-range dependencies, suffers from poor fusion quality of heterogeneous features, and exhibits limited model interpretability—thereby hindering practical deployment in instructional intervention contexts. To address these issues, a unified framework for behavioral analysis and engagement prediction was proposed. Relying on non-intrusive sensing techniques, multimodal learning behavior data were collected, and learning engagement was quantified through the fusion of subjective and objective evaluation metrics. Temporal behavioral topological associations were uncovered by constructing horizontal visibility graphs integrated with a graph attention network. A cross-modal attention mechanism was employed to adaptively fuse multi-source heterogeneous features, while multitask learning was introduced to enhance model generalization, enabling accurate prediction of learning engagement trajectories. Furthermore, a gradient-weighted attribution strategy was incorporated to identify key influencing factors, and a mobile intervention mechanism was designed to optimize learners’ behavioral states. This study provides a technical reference for mobile learning behavior mining, learning state prediction, and the development of intelligent educational applications.