基于超图神经网络以多角度概念特征融合的概念先决学习

Concept Prerequisite Learning with Multi-View Concept Feature Fusion Based on Hypergraph Neural Network

概念之间的先决条件关系是智慧教育领域开展个性化学习相关工作的基础性任务,具有至关重要的作用。现有研究中基于特征计算的方法依赖于手工特征提取,受限于本文结构,基于二元图结构方法则忽略了概念和文档对象两两之间的复杂高阶关系。为了解决以上问题,该文提出HyperCPRL,基于超图编码高阶拓扑结构的能力,从三个角度构造基于概念结构、概念语义距离和文档-概念隶属关系三个超图,以捕捉建模对象之间具有的复杂关系特征,再融合三个超图结点表征,并利用自注意力机制在概念全域进一步挖掘先决关系,利用孪生网络实现先决关系预测。在四个真实数据集上进行了大量实验对比,HyperCPRL取得了较好的效果,且对包含低度概念样本的识别能力更强。

Prerequisite relations among concepts play a crucial role as a foundational task related to individualized learning in wisdom education field.In the existing research,the feature-based methods depend on manual features extraction,while the binary graph-based methods ignore the complex high-order relations between concepts and documents.The HyperCPRL is proposed to solve the above problems,which uses hypergraph to encode high-order topology.The model construct three hypergraphs based on conceptual structure,conceptual semantic distance and document concept membership relationship from three perspectives.Then,the model fuse the representations of three hypergraph nodes,and use the self-attention mechanism to further mine the relations in the whole concept domain.At last,siamese network is used to predict prerequisite relations.Extensive experiments on four datasets demonstrate the efficacy of HyperCPRL for its recognition ability for samples with low-degree concepts.