基于异构网络表征学习的作者学术行为预测

Author’s Academic Behavior Prediction Based on Heterogeneous Network Representation Learning

作者学术行为预测旨在从异构学术网络中挖掘作者的行为关系,以促进科研合作,产出高水平、高质量的研究成果。现有的节点表示方法大多未考虑节点的语义特征、内容特征、全局结构等,难以有效学习网络中节点的低维特性。为有效融合节点的多维特征和全局结构,提出了一种集成BiLSTM、注意力机制和聚类算法的异构网络表示学习方法HNEMA,以提高学术网络中作者的学术行为预测效果。HNEMA首先基于BiLSTM和注意力机制融合节点的多维特征,聚合同一元路径下或不同元路径下相同类型的邻居,随后聚合待表征节点的所有邻居的多维特征。基于此,采用聚类算法捕获节点的全局特征,从而全面有效地学习节点的低维特性。在全面特征学习的基础上,应用逻辑回归分类器预测作者的学术行为。在3个公开数据集上的验证实验结果表明,相比其他方法,HNEMA在AUC和F1指标上都有一定程度的提升。

The author’s academic behavior prediction aims to mine the behavioral relationships of authors from heterogeneous academic networks to promote scientific research cooperation and produce high-level and high-quality research results.Most of the existing methods of node representation learning do not consider the semantic feature, content feature, global structure of the node, etc.It is difficult to effectively learn the low-dimensional characteristics of the node in the network.In order to effectively integrate the multi-dimensional features and global structure of nodes, a heterogeneous network representation learning method(HNEMA) that integrates BiLSTM,attention mechanism and clustering algorithm is proposed to improve the predictive effect of author’s academic behavior.HNEMA first integrates the multi-dimensional features of nodes based on BiLSTM and attention mechanism, aggregates the same type of neighbors on the same meta-path or different meta-paths, and then aggregates the multi-dimensional features of all neighbors of the node to be characterized.Based on this, a clustering algorithm is used to capture the global features of the node, so as to comprehensively and effectively learn the low-dimensional characteristics of the node.On the basis of comprehensive feature learning, logistic regression classifier is used to predict author’s academic behavior.Validation experiments on three public datasets show that HNEMA has a certain degree of improvement in AUC and F1 indicators compared to other methods.