Research on student academic-psychological joint early warning model under multi-task deep learning framework

Early identification of students at risk is critical for modern educational management. However, existing systems often treat academic performance and psychological well-being as independent domains, neglecting their complex interdependencies and shared behavioral markers. This paper proposes a novel multi-task learning (MTL) framework designed for joint academic and psychological early warning. By integrating multi-source heterogeneous data—ranging from temporal behavioral logs, such as attendance and campus consumption, to unstructured counseling records processed via BERT-based semantic embeddings—the framework captures a holistic view of student dynamics. To address the "seesaw effect" inherent in MTL, we employ a Multi-gate Mixture-of-Experts (MMoE) architecture that effectively decouples task-specific features while maintaining a shared knowledge base. Furthermore, we introduce an adaptive gradient normalization (GradNorm) mechanism to dynamically balance the loss weights between the academic GPA regression task and the psychological crisis classification task. This prevents gradient dominance from denser data sources and ensures stable convergence across disparate objectives. Experimental results demonstrate that the proposed approach achieves significant performance gains, particularly in the recall of psychological crisis detection, outperforming several state-of-the-art baselines. The proposed framework advances artificial intelligence in education by integrating multimodal student data with multi-task deep learning for joint academic prediction and psychological-risk early warning. Its interpretable and fairness-aware design demonstrates practical AI applications supporting timely, data-driven educational decision-making. The framework is deployed through a real-time inference pipeline powered by Apache Flink to ensure campus-wide scalability and low-latency response. Finally, the integration of SHAP values enhances model transparency, providing educators with actionable insights and interpretable evidence for targeted interventions. This research offers a robust, scalable, and interpretable solution for proactive student support in large-scale academic environments.