Scholarlink

<p>The scalability of massive open online courses (MOOCs) is hampered by significant challenges for recommendation systems, which must operate under extreme data sparsity, rapid concept drift, and stringent low-latency constraints. To address this, we propose a novel hybrid architecture that synergistically integrates stochastic frequent-pattern mining (SFPM) utilizing Lossy Counting and Count-Min Sketch with a transformer-based deep sequence model. The SFPM layer maintains a real-time, approximate model of frequent sequential patterns across the entire user population, providing a continuous stream of recency-adjusted, statistically grounded features (e.g., support, confidence). These global features are dynamically infused into the deep neural predictor, which concurrently processes individual learner trajectories to generate personalized recommendations. We present a formal complexity and latency analysis, confirming the model’s operational efficiency. Comprehensive evaluations on two public educational datasets (EdNet and ASSISTments) demonstrate that our approach outperforms state-of-the-art sequential recommenders (e.g., SASRec, BERT4Rec) and collaborative filtering baselines, achieving an average improvement of 9.5% in NDCG@10 and 11.2% in HR@10 while reducing the 95th percentile inference latency by 43%. Furthermore, the model exhibits enhanced robustness to concept drift and provides interpretable, pattern-level attributions, offering actionable insights for educational stakeholders.</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

SPARL: a hybrid architecture for real-time educational recommendation via streaming pattern mining and sequential learning

  • 期刊论文

  • 发表日期：

    2025年11月3日

查看全文
  • S Basheer Ahamed,
  • A. R. Mohamed Shanavas

摘要

The scalability of massive open online courses (MOOCs) is hampered by significant challenges for recommendation systems, which must operate under extreme data sparsity, rapid concept drift, and stringent low-latency constraints. To address this, we propose a novel hybrid architecture that synergistically integrates stochastic frequent-pattern mining (SFPM) utilizing Lossy Counting and Count-Min Sketch with a transformer-based deep sequence model. The SFPM layer maintains a real-time, approximate model of frequent sequential patterns across the entire user population, providing a continuous stream of recency-adjusted, statistically grounded features (e.g., support, confidence). These global features are dynamically infused into the deep neural predictor, which concurrently processes individual learner trajectories to generate personalized recommendations. We present a formal complexity and latency analysis, confirming the model’s operational efficiency. Comprehensive evaluations on two public educational datasets (EdNet and ASSISTments) demonstrate that our approach outperforms state-of-the-art sequential recommenders (e.g., SASRec, BERT4Rec) and collaborative filtering baselines, achieving an average improvement of 9.5% in NDCG@10 and 11.2% in HR@10 while reducing the 95th percentile inference latency by 43%. Furthermore, the model exhibits enhanced robustness to concept drift and provides interpretable, pattern-level attributions, offering actionable insights for educational stakeholders.