基于图神经网络异构数据融合的学科新兴主题探测研究

Detecting Scientific Emergency Topic Based on Heterogeneous Data Fusion Using GCN

[研究目的]数据异构性阻碍了大数据集成分析,而异构数据的深度融合学习能够增强学科数据分析能力,为预见学科新兴主题提供有力支撑。[研究方法]探测分析由两部分衔接构成,一是实现多元异构学科数据深度融合的图卷积神经网络(GCN),二是旨在学科主题预测的LSTM模型。具体地,通过GCN的深度学习能力,将包含多维特征和共现关系的异构主题数据转化为同构表示向量,不但实现异构融合,更为后续预测模型提供统一数据基础;然后,将主题表示向量时间序列输入LSTM模型,预测学科主题的新兴特征,为前瞻预见学科新兴主题提供决策支持。[研究结论]以图书情报学为对象的实证充分检验了GCN+LSTM的设计合理性,融合模型比非融合模型在主题趋势预测中展现出明显优势。

[Research purpose]Data heterogeneity makes large data integration analysis difficult.Deep fusion learning for data with various structures aids in improving academic data analysis capability,and support the prediction of scientific emergency topics.[Research method]Two components make up detection analysis:(1)Graph Convolution Network(GCN)for deep fusion with various and heterogeneous academic data.(2)LSTM model for topic prediction in academic fields.In particular,using deep learning capability of GCN,heterogeneous topics data,including multi-characteristics and co-occurrence relations,are transformed into homogeneous representation vectors,realizing heterogeneous fusion while also providing a unified data base for the subsequent prediction model.In order to anticipate the emergency characteristics of academic topics and provide decision assistance for predicting academic emergency topics,topic representation vectors are then fed into a LSTM model to predict academic emergency characteristics,giving decision assistance for predicting academic emergency topics.[Research conclusion]In the academic discipline of library and information science,the empirical findings support the design of GCN+LSTM model as being reasonable.In addition,the fusion model outperformed than non-fusion models.