一种基于BiLSTM的混合层次化图分类模型
A Hybrid Hierarchical Graph Classification Model Based on BiLSTM
张红梅 1郑创 2钟晓雄3
作者信息
- 1. 桂林电子科技大学信息与通信学院,广西 桂林 541004;鹏程实验室,广东 深圳 518055
- 2. 桂林电子科技大学信息与通信学院,广西 桂林 541004
- 3. 鹏程实验室,广东 深圳 518055
- 折叠
摘要
图分类在化学和生物信息学等诸多领域中是一个非常重要且极具挑战的问题,GNN模型是图分类问题的主流方法.现有的GNN模型采用卷积操作来实现邻域节点信息聚集,再通过池化操作生成粗化图.然而,仅通过池化方法不能捕获到每次卷积后读出图的双向依赖关系.为了提取到更充分的特征信息,提出一种混合层次化模型,首先分别提取节点特征信息和结构特征信息,再将特征信息融合,然后采用BiLSTM捕获不同层次读出图之间的双向依赖关系,从而提取到更丰富的特征信息.实验结果表明,与对比模型相比,上述模型的准确度有着明显的提升.
Abstract
Graph classification is a very important and challenging problem in many fields such as chemistry and bioinformatics,and the GNN model is the mainstream method for graph classification problems.The existing GNN models use convolution operations to aggregate the information of neighbor nodes,and then generate the coarsening-graphs by pooling method.However,the bidirectional dependencies of the readout graph after each convolution cannot be captured by the pooling method alone.In order to extract more sufficient feature information,a hybrid hierarchical model is proposed in this paper,which first extracts node feature information and structural feature information respec-tively,then fuses the feature information,and then uses BiLSTM to capture the bidirectional dependencies between different levels of readout graphs,so as to extract richer feature information.The experimental results show that com-pared with the comparison model,the accuracy of the model has been significantly improved
关键词
图神经网络/层次顺序/双向依赖关系Key words
Graph neural network/Hierarchical order/Bidirectional dependencies引用本文复制引用
基金项目
广西自然科学基金重点项目(2020GXNSFDA238001)
鹏程实验室重大攻关项目(PCL2021A02)
桂林电子科技大学研究生教育创新项目(2021YCXS046)
出版年
2024
NETL
NSTL
万方数据