科研信息门户的资源推荐技术研究

Research on Resource Recommendation Technology of Scientific Research Information Portal

【应用背景】科研信息门户作为科研工作者获取资源服务的入口,已经成为了科研人员、管理决策者、学生等多种用户的工作台,在科研活动、科研管理、教育培训、科学传播等多个业务领域发挥着重要作用。【目的】针对科研信息门户中存在的信息资源配置不合理问题,设计出适用于科研信息门户的推荐算法来提高信息资源的推送效率,对于科研工作者是十分有意义的。【方法】本文提出了一个混合的推荐算法,对于首次使用系统的新用户,可以基于用户属性,通过K-means聚类后找到相邻用户来计算预测评分,对于存在行为数据的用户,先通过计算用户与资源的相似度来解决隐式反馈缺少负反馈的问题,再使用矩阵分解的方法计算预测评分。最后将两种算法的预测评分进行线性组合得到最终预测评分,该算法既利用了群体智慧也体现了个性化。【结果】通过在真实的科研信息门户网站上进行代码埋点来采集用户行为数据,完成对比试验,证明提出的推荐方法能在解决冷启动问题的同时保证较高的推荐准确率。

[Application Background]As the entrance for scientific researchers to obtain resource services,the scientific research information portal has become a workbench for scientific researchers,management decision makers,students and other users.It is used in scientific research activities,scientific research management,education and training,and scientific communication.Each business area plays an important role.[Objective]To address the issue of unreasonable information resources allocation in scientific research information portals,design of recommendation technology suitable for scientific research information portals to improve the push efficiency of information resources is of great significance to scientific researchers.[Methods]This paper proposes a hybrid recommendation algorithm.For new users who use the system for the first time,based on user attributes,neighbor users can be found through K-means clustering to calculate prediction scores.For old users,the problem of lacking negative feedback in implicit feedback is firstly solved by calculating the similarity between users and resources,and then matrix factorization is used to calculate the predicted score.Finally,the prediction scores of the two algorithms can be linearly combined to obtain the final prediction score.The algorithm not only employs the wisdom of the group but also embodies personalization.[Results]Based on user behavior data by embedding code on the real scientific research information portal website,the comparative experiment proves that the proposed recommendation method can solve the cold start problem while ensuring high recommendation accuracy.