基于分子图相似度的医药文献推荐方法  

A Recommender via Similarity of Molecule Graphs for Medical Literature

作　　者：冯贤兵 陶涛[1] 吕肖庆[1] 

机构地区：[1]北京大学王选计算机研究所,北京

出　　处：《计算机科学与应用》2022年第12期2853-2862,共10页Computer Science and Application

摘　　要：当今生物医学等领域的文献快速增长,一方面促进了科研交流,但同时也为研究人员带来了巨大的阅读压力,尽管业界已出现了一些论文搜索和推荐的方法,但其大都只依据论文的元信息和文本信息,而对文章内容,尤其是插图等非文字对象尚未充分挖掘并利用,因此现有系统在给读者的推荐结果中,还存在着大量重复、泛化等低效情况。为此,我们探索并建立了一个基于论文内容的文档级推荐系统,具体包括:文档解析、文本对象理解、内容相似性度量、多级索引机制、以及优化推荐结果等主要环节。其中,针对生物医学类科技文献中特有的分子式图片,我们提出了一种图相似度的度量方法,即半分支编辑距离(Half-branch GED,简称HB-GED)算法,同时针对分子图形表示和文档之间关系表示也提出了图卷积模型。在真实数据集上的实验结果表明,本文提出的论文推荐方法,可有效筛选出更符合查询者意图的候选论文。Nowadays, the consistent growth of scientific and technical literature leads to formidable pressure on medical researchers. Researchers turn to the search engine and paper recommender systems and still have to spend more time keeping up with the trends and directions in their field. However, most existing recommender approaches mainly depend on text-based information and ignore non-text objects, such as informative figures. To this end, we establish a document-to-document recommender system for medical literature. Specifically, we proposed a deep-learning-based seg-mentation method for extracting molecular graphs, a Half-branch GED algorithm for evaluating the similarity of molecules, and a bipartite-graph-based algorithm for paper similarity, respectively. Experimental results on real-world datasets demonstrate the effectiveness of the proposed recom-mender system.

关 键 词：图相似度 分子图 论文推荐 图编辑距离 二部图 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术]