Scholarlink

<p>This study proposes a high-precision MOOC recommendation algorithm that integrates deep learning with an enhanced self-attention mechanism, designed to operate efficiently under large-scale data scenarios. Leveraging the MOOCCubeX dataset, which contains over 296 million learner–course interaction records, the computational demand for model training is substantial due to the quadratic complexity of the self-attention mechanism. To address this, we optimize the SASRec architecture through three key enhancements: an adaptive sequence masking mechanism to capture temporal intervals, the AdamW optimizer for stability, and a class-imbalance-aware loss function. Furthermore, we adopt a resource-efficient training workflow using GPU-accelerated parallel computing to ensure feasibility on standard hardware. Experimental results demonstrate that our optimized model achieves significant performance gains (HR@10 = 0.897, NDCG@10 = 0.658) while maintaining efficient large-batch training. The results highlight that algorithmic innovations combined with hardware-aware optimization are essential to support scalable sequential recommendation models in real-time educational applications.</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

Research on precision recommendation algorithm based on the integration of deep learning and self-attention mechanism

  • 期刊论文

  • 发表日期：

    2026年1月9日

查看全文
  • Wenxin Zhao,
  • Hang Su,
  • Zhongjian Wang,
  • Shuo Xu,
  • Lei Zhao,
  • Zhenbin Liu

摘要

This study proposes a high-precision MOOC recommendation algorithm that integrates deep learning with an enhanced self-attention mechanism, designed to operate efficiently under large-scale data scenarios. Leveraging the MOOCCubeX dataset, which contains over 296 million learner–course interaction records, the computational demand for model training is substantial due to the quadratic complexity of the self-attention mechanism. To address this, we optimize the SASRec architecture through three key enhancements: an adaptive sequence masking mechanism to capture temporal intervals, the AdamW optimizer for stability, and a class-imbalance-aware loss function. Furthermore, we adopt a resource-efficient training workflow using GPU-accelerated parallel computing to ensure feasibility on standard hardware. Experimental results demonstrate that our optimized model achieves significant performance gains (HR@10 = 0.897, NDCG@10 = 0.658) while maintaining efficient large-batch training. The results highlight that algorithmic innovations combined with hardware-aware optimization are essential to support scalable sequential recommendation models in real-time educational applications.