结合链路预测和ET机器学习的科研合作推荐方法研究

发布时间：2018-07-22 15:07

【摘要】：【目的】结合链路预测与机器学习,提出推荐未来科研合作的新方法,以提高单独基于链路预测方法的推荐精确度。【方法】构建加权作者合作网,以不同的链路预测指标作为特征输入,运用极端随机树(Extremely Randomized Trees,ET)机器学习算法训练分类,并利用遍历算法求取分类结果的最优权重组合,选取TOP准确度的预测作为合作推荐结果。【结果】选取纳米科技领域2008年 2010年SCI论文数据进行实证。在城市合作推荐中,改进的ET方法优于已有方法,有良好的推荐成功率;预测方法受网络结构等因素影响较小,适用范围更广泛。【局限】科研合作受合作动机、地域、语言等诸多因素影响,加权作者合作网没有反映在一篇论文中同城市、同机构的多个作者,也没有反映上述因素。【结论】改进算法能够比单个预测指标产生更准确的合作推荐建议,也为推广到大学等机构、个人等更微观的应用层面提供参考。
[Abstract]:[objective] combined with link prediction and machine learning, a new method for recommending future scientific research cooperation is proposed to improve the recommendation accuracy based on link prediction alone. [methods] A weighted author cooperative network is constructed. With different link prediction indexes as feature input, the classification is trained by extreme random trees (Extremely randomly trees) machine learning algorithm, and the optimal weight combination of classification results is obtained by traversal algorithm. The prediction of top accuracy was selected as the cooperative recommendation result. [results] the data of sci in the field of nanotechnology in 2008 and 2010 were selected for empirical study. In the urban cooperative recommendation, the improved et method is superior to the existing method and has a good success rate of recommendation, the prediction method is less affected by the network structure and other factors, and the scope of application is wider. Many factors, such as language, are not reflected in a single paper by multiple authors in the same city and institution. [conclusion] the improved algorithm can produce more accurate recommendations for cooperation than a single prediction index, and can also provide a reference for the extension to institutions such as universities, individuals and other more micro application levels.
【作者单位】：中国科学院文献情报中心;中国科学院大学;中国科学院科技战略咨询研究院;
【基金】：国家自然科学基金面上项目“科学结构特征及其演化动力学分析方法与应用研究”(项目编号:71173211)的研究成果之一
【分类号】：G353.1

【参考文献】