Scholarlink

<p>To address the dilemma of homogeneous talent training and the efficiency bottleneck of human resource management in universities, this study proposes an innovative personalized training framework integrating artificial intelligence, big data, and deep learning. Based on the 18-dimensional full-cycle behavior dataset of 5,000 students and OULAD dataset, a multimodal heterogeneous data fusion pipeline is constructed. This study adopts Generative Adversarial Network (GAN) for data imputation and bias optimization, designs Hierarchical Attention Graph Neural Network (HA-GNN) to capture hierarchical correlations among features, and uses Long Short-Term Memory (LSTM) to model temporal behavior patterns. The experimental results demonstrate that, under 10 independent repeated runs with random seed variation, the Hierarchical Attention Graph Neural Network-Long Short-Term Memory (HA-GNN-LSTM) model achieves lower prediction error on the academic performance prediction task, with a Mean Absolute Error (MAE) of 4.2 ± 0.3. Compared with the Temporal Fusion Transformer (TFT) baseline model, MAE is reduced by 31.1%. Welch’s two-tailed t-tests based on independent run results remain statistically significant after Holm-Bonferroni multiple comparison correction <InlineEquation ID="IEq1"><EquationSource Format="TEX">\(\:\left({p}_{\text{a}\text{d}\text{j}}&lt;0.01\right)\)</EquationSource></InlineEquation>. The Normalized Discontinued Cumulative Gain at Top 5 (NDCG @ 5) index of personalized recommendation system reaches 0.90, which verifies the effectiveness of spatio-temporal feature modeling. At the management application level, the improvements in advisor allocation response time and resource idle rate are derived from simulation experiments based on historical data replay, rather than online deployment in real campus management systems. The simulation results demonstrate that, under established constraints and historical sample distributions, advisor allocation response time could be reduced by 60% and resource idle rate could be decreased by 63.4%. These findings indicate the framework’s potential for optimizing educational resource allocation. However, its managerial benefits require further validation through subsequent real-world deployment and long-term follow-up studies.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Personalized talent cultivation and academic prediction framework for higher education based on the HA-GNN-LSTM architecture

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Qi Gong,
  • Jing Shi

摘要

To address the dilemma of homogeneous talent training and the efficiency bottleneck of human resource management in universities, this study proposes an innovative personalized training framework integrating artificial intelligence, big data, and deep learning. Based on the 18-dimensional full-cycle behavior dataset of 5,000 students and OULAD dataset, a multimodal heterogeneous data fusion pipeline is constructed. This study adopts Generative Adversarial Network (GAN) for data imputation and bias optimization, designs Hierarchical Attention Graph Neural Network (HA-GNN) to capture hierarchical correlations among features, and uses Long Short-Term Memory (LSTM) to model temporal behavior patterns. The experimental results demonstrate that, under 10 independent repeated runs with random seed variation, the Hierarchical Attention Graph Neural Network-Long Short-Term Memory (HA-GNN-LSTM) model achieves lower prediction error on the academic performance prediction task, with a Mean Absolute Error (MAE) of 4.2 ± 0.3. Compared with the Temporal Fusion Transformer (TFT) baseline model, MAE is reduced by 31.1%. Welch’s two-tailed t-tests based on independent run results remain statistically significant after Holm-Bonferroni multiple comparison correction \(\:\left({p}_{\text{a}\text{d}\text{j}}<0.01\right)\). The Normalized Discontinued Cumulative Gain at Top 5 (NDCG @ 5) index of personalized recommendation system reaches 0.90, which verifies the effectiveness of spatio-temporal feature modeling. At the management application level, the improvements in advisor allocation response time and resource idle rate are derived from simulation experiments based on historical data replay, rather than online deployment in real campus management systems. The simulation results demonstrate that, under established constraints and historical sample distributions, advisor allocation response time could be reduced by 60% and resource idle rate could be decreased by 63.4%. These findings indicate the framework’s potential for optimizing educational resource allocation. However, its managerial benefits require further validation through subsequent real-world deployment and long-term follow-up studies.