Multi-source behavioral data and deep learning model for student performance prediction in vocational education

Vocational education increasingly adopts blended learning approaches, combining online and traditional instruction to enhance skill development and learning outcomes. As students engage in various digital activities, they generate large volumes of behavioral data that can be leveraged to predict academic performance and enable early intervention for at-risk learners. This research proposes a deep learning-based student performance prediction framework tailored to vocational education environments, utilizing multi-source behavioral data. Data were collected from 2000 vocational learners enrolled in a semester-long program. The framework utilizes multi-source behavioral data, including Learning Management System (LMS) activity logs, internet usage patterns, attendance records, assessments, and demographic information, to enhance predictive accuracy and support early intervention in vocational education. Data were preprocessed by handling missing values, data cleaning, and min-max normalization. Principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) were fused to extract dynamic behavioral features. The proposed model, the Seven-spot Ladybird Optimized Multi-Source Attention-based Gated Recurrent Units (SLO-MSAGRU) model, is applied to predict student performance in vocational education. Implemented in Python, the findings show that the SLO-MSAGRU approach performs better than multimodal baseline architectures, achieving superior results with accuracy, F1-score, recall, and precision ranging from 95% to 98%. These findings reveal that broader internet behavior patterns contribute more significantly to student performance prediction than platform-specific learning data. The framework offers a practical tool for educators to apply personalized learning strategies and timely interventions, thereby improving decision-making in vocational training environments.