Scholarlink

<p>Drug recommendation aims to recommend proper drugs according to patients’ symptoms. A patient usually has multiple symptoms, and a doctor needs to prescribe a combination of drugs rather than an individual drug for a patient. However, existing works mainly model the semantic interactions between symptoms and drugs at the item level but ignore set-level semantic interactions, which are also crucial for drug recommendation tasks. To address the above issues, we propose <b>Cross</b>-view <b>G</b>raph <b>C</b>ontrastive <b>L</b>earning with Dual Tasks for Drug Recommendation (<b>CrossGCL</b>) to recommend drugs. Our model learns the representation of symptoms and drugs from the item-level interactions (item view) and the set-level interactions (set view) and then maximizes the mutual information from the two views by cross-view contrastive learning so that the representation from different views can be enhanced by each other. Besides, we construct two structural-duality learning tasks, a symptom-driven drug prediction task and a drug-driven symptom prediction task, to better capture the set-level and item-level semantic information. We conduct experiments on the widely used public MIMIC-III dataset and show that our model outperforms state-of-the-art models.</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

CrossGCL: cross-view graph contrastive learning with dual tasks for drug recommendation

  • 期刊论文

  • 发表日期：

    2025年8月20日

查看全文
  • Wushao Wen,
  • Liang Wang,
  • Lihuang Fang,
  • Qiangpu Chen,
  • Jinghui Qin

摘要

Drug recommendation aims to recommend proper drugs according to patients’ symptoms. A patient usually has multiple symptoms, and a doctor needs to prescribe a combination of drugs rather than an individual drug for a patient. However, existing works mainly model the semantic interactions between symptoms and drugs at the item level but ignore set-level semantic interactions, which are also crucial for drug recommendation tasks. To address the above issues, we propose Cross-view Graph Contrastive Learning with Dual Tasks for Drug Recommendation (CrossGCL) to recommend drugs. Our model learns the representation of symptoms and drugs from the item-level interactions (item view) and the set-level interactions (set view) and then maximizes the mutual information from the two views by cross-view contrastive learning so that the representation from different views can be enhanced by each other. Besides, we construct two structural-duality learning tasks, a symptom-driven drug prediction task and a drug-driven symptom prediction task, to better capture the set-level and item-level semantic information. We conduct experiments on the widely used public MIMIC-III dataset and show that our model outperforms state-of-the-art models.