Comparative experimental analysis.

Student well-being prediction is of great significance for promoting personalized education and preventing mental health problems, but existing methods suffer from limitations including lack of psychological theory guidance, neglect of student relationship modeling, and insufficient cross-cultural adaptability. This study proposes the PERMA-GNN-Transformer model, which innovatively integrates Seligman’s PERMA positive psychology theory with graph neural networks and Transformer architecture. The model achieves the transformation from raw educational data to psychologically meaningful representations through a theory-driven feature embedding mechanism, designs four types of student relationship graphs based on cosine similarity, Euclidean distance, learning styles, and PERMA weighting, and employs a five-head attention mechanism corresponding to the five PERMA dimensions. Experiments on the Western cultural background Lifestyle and Wellbeing Data (n = 12,757) and the East Asian cultural background International Student Mental Health Dataset (n = 268) demonstrate that compared to the optimal baseline methods, our proposed model achieves an 18.9% performance improvement on large-scale datasets and a 27.8% improvement on small-scale datasets, with the PERMA comprehensive evaluation metric reaching 0.792 and passing statistical significance tests at p学生幸福感预测对于推进个性化教育实践、防范心理健康问题具有重要价值，但现有方法存在诸多局限：缺乏心理学理论指导、忽视学生关系建模，且跨文化适配性不足。本研究提出PERMA-GNN-Transformer模型，创新性地将塞利格曼的PERMA积极心理学理论（PERMA）与图神经网络（Graph Neural Networks, GNN）及Transformer架构相结合。该模型通过理论驱动的特征嵌入机制，实现了从原始教育数据到具有心理学意义的表征的转换；设计了基于余弦相似度、欧氏距离、学习风格与PERMA权重的四类学生关系图，并采用了对应PERMA五大维度的五头注意力机制。在西方文化背景下的《生活方式与幸福感数据集》（Lifestyle and Wellbeing Data，n=12,757）以及东亚文化背景下的《国际学生心理健康数据集》（International Student Mental Health Dataset，n=268）上开展的实验表明，相较于最优基线方法，本研究提出的模型在大规模数据集上性能提升18.9%，在小规模数据集上提升27.8%，其PERMA综合评价指标达到0.792，并通过了p值的统计显著性检验。