基于LSTM模型与加权链路预测的学科新兴主题成长性识别研究  

Identifying Growth of Discipline Topics Using LSTM and Weighted Link Prediction

作　　者：段庆锋 陈红 刘东霞 闫绪娴[1] 张红兵[1] Duan Qingfeng;Chen Hong;Liu Dongxia;Yan Xuxian;Zhang Hongbing(School of Management,Shanxi University of Finance&Economics,Taiyuan 030006,China)

机构地区：[1]山西财经大学管理科学与工程学院,山西太原030006

出　　处：《现代情报》2022年第9期37-48,142,共13页Journal of Modern Information

基　　金：教育部人文社会科学项目“基于学术社交媒体的学科新兴趋势识别研究”(项目编号:20YJA870005);教育部人文社会科学项目“学术资源配置公平、效率与影响因素研究:学者、大学与区域的多层嵌入”(项目编号:19YJAZH052);国家社会科学基金项目“供给侧改革背景下提升资源型企业经济韧性的关键要素、效应评估与实现路径研究”(项目编号:20BGL100);山西省高等学校教学改革创新项目“面向财经院校的创业基础课程建设研究”(项目编号:J2019107);山西省研究生教改研究课题“‘新文科’背景下财经院校研究生双创教育模式研究”(项目编号:2019011)。

摘　　要：[目的/意义]揭示学科主题潜在的成长性趋势是识别其新兴特征的关键和难点。[方法/过程]从热度和影响力两个维度考察主题成长性,通过基于机器学习算法的预测模型估计主题潜在成长性,基于此形成新兴主题的细分类型划分。首先,设计融合文献计量指标和替代计量指标的主题热度指标,构建基于长短记忆神经网络LSTM的主题热度预测模型,预测主题热度增长率;其次,基于加权链路预测相似性指标,构建旨在预测未来主题网络的三层神经网络预测模型,并采用PageRank算法得到主题影响力增长率预测值;最后,基于热度增长率预测值和影响力增长率预测值构建二维识别空间,通过主题聚类,识别不同子类型的学科新兴主题。[结果/结论]以情报学为领域的实证研究检验了识别方法的有效性,反映了成长性预测指标对于新兴特征的敏感捕捉能力。[Purposes/Significance]The key and difficult point during the period when identifying emergent characteristics of topic in discipline is to disclose growing trend in potentially.[Method/Process]The topic’s growth was investigated from two aspects of hotness and influence, and the extent to which topics grow up was estimated by using machine-learning algorithm as predicting model.Based on this, emerging topics were classified into different sub-categories.Firstly, hotness index that combines bibliometric indicator and altmetric indicator was designed, and then a prediction model that use LSTM to predict the extent to which hotness index increase was established.Secondly, another prediction model that has three-layer neural network and can predict newly occurred link in future between two topics based on the similarity index from weighted link prediction was proposed.Based on those, PageRank algorithm was used to estimate influence of a topic embedded in the network we had predicted.Finally, a comprehensive method was offered to discern different types of emergent topic.The method constructed a two-dimensional recognition space, using the growth indicators including hotness and influence, to conduct clustering analysis on topics.[Result/Conclusion]The paper conducted an empirical study with the samples from the discipline of information science, which successfully confirmed the effectiveness of our proposed method.Results illustrated that those index for growth prediction are sensitive enough to emergency.

关 键 词：学科新兴主题 成长性 识别 加权链路预测 LSTM 

分 类 号：G254[文化科学—图书馆学]