基于动态多任务学习的科技文献推荐模型构建及实证研究  

Construction and Empirical Study of Scientific and Technological Literature Recommendation Model Based on Dynamic Multi-task Learning

作　　者：李洁[1,2] 张国标 周毅 郗玉娟[5] 杨金庆 Li Jie;Zhang Guobiao;Zhou Yi;Xi Yujuan;Yang Jinqing(Institute of Intelligent Society and Data Governance,Soochow University,Suzhou 215000;Shandong University Library,Jinan 250100;School of Communication,Soochow University,Suzhou 215000;School of Sociology,Soochow University,Suzhou 215000;School of Innovation and Entrepreneurship,Shandong University,Qingdao 266100;School of Information Management,Central China Normal University,Wuhan 430079)

机构地区：[1]苏州大学智能社会与数据治理研究院,苏州215000 [2]山东大学图书馆,济南250100 [3]苏州大学传媒学院,苏州215000 [4]苏州大学社会学院,苏州215000 [5]山东大学国际创新转化学院,青岛266100 [6]华中师范大学信息管理学院,武汉430079

出　　处：《图书情报工作》2024年第13期122-131,共10页Library and Information Service

基　　金：江苏省社会科学青年基金项目“疫情常态化背景下图书馆数字资源认知推荐研究”(项目编号:21TQC001);国家社会科学基金青年项目“模糊认知视角下智慧图书馆资源推荐服务模式及实证研究”(项目编号:22CTQ009);湖北省自然科学基金面上项目“基于大语言模型深度语义理解的领域技术谱系生成研究”(项目编号:2024AFB1018)研究成果之一。

摘　　要：[目的/意义]为实现科技文献推荐场景要素的交互增强,将各要素交互特性捕捉问题转化为多任务共同优化学习问题,构建基于动态多任务学习的科技文献推荐模型,以进一步提升科技文献推荐性能。[方法/过程]采用多任务学习方法,针对科技文献推荐要素可采集的关键特征进行子任务解构,借助多头注意力机制,进行子任务交互关系的动态学习,在动态学习各任务交互关系的基础上设计科技文献推荐模型。[结果/结论]根据CiteULike数据实验结果,所构建的DMRSTL模型在3个评价指标上均显著优于对比模型,最高差值为AUC指标提升15.51%,MRR指标提升11.90%,nDCG@5指标提升16.45%,且通过任务组合对比实验进一步表明,借助推荐要素的交互增强,可以有效提升科技文献的推荐性能。[Purpose/Significance]In order to enhance the interaction of scientific and technological literature recommendation scene elements and transform the problem of capturing the interactive characteristics of each element into a multi-task joint optimization learning problem,this paper constructs a scientific and technological literature recommendation model based on dynamic multi-task learning to further improve the performance of scientific and technological literature recommendation.[Method/Process]Based on multi-task learning method,the sub-tasks are deconstructed according to the key features collected from scientific and technological literature recommendation elements,and the multi-head attention mechanism was used to dynamically learn the interactive relationships of sub-tasks.A scientific and technological literature recommendation model was designed through dynamic learning of the interaction of each task.[Result/Conclusion]According to the experimental results of CiteULike data,the DMRSTL model constructed in this article is significantly better than the comparison model in three evaluation indicators.The highest difference is the increase of AUC indicator by 15.51%,MRR by 11.90%,and nDCG@5 indicator by 16.45%.The task combination comparative experiments further show that the interactive enhancement of recommendation elements can effectively improve the recommendation performance of scientific and technological literature.

关 键 词：科技文献推荐 多任务学习 多头注意力机制 任务交互 

分 类 号：G258.6[文化科学—图书馆学] TP391.3[自动化与计算机技术—计算机应用技术]