基于可微池化的层级图相似性学习
Hierarchical graph similarity learning based on differentiable pooling
吴磊 1李晓楠 1李冠宇1
作者信息
- 1. 大连海事大学信息科学技术学院,辽宁大连 116026
- 折叠
摘要
目前,大多数关于图相似性学习的工作仅考虑图级匹配或节点级匹配,忽略了多层级的粗图级匹配,为解决该问题提出一种可微池化层级图匹配网络(PHMN)模型.逐层将图的节点表示进行软聚类进而将源图转化为尺寸缩小的粗图;在图对上使用多角度多层级的跨图匹配层,获取匹配矩阵;由注意力机制将图对匹配矩阵转化为匹配向量后,传入LSTM模型和多层感知机进行相似度预测.该模型在图回归任务和图分类任务的对比实验中,分别取得8项最优表现和6项最优表现.
Abstract
Currently,most studies on graph similarity learning only consider graph level matching or node level matching,but ignore the multi-level rough graph matching.To solve this problem,a DiffPool hierarchical graph matching network(PHMN)model was proposed.The source graph was transformed into a coarse graph with reduced size by soft clustering of graph node representation layer by layer.A multi-angle and multi-level cross-graph matching layer was used to obtain the matching matrix.The graph pair matching matrix was transformed into matching vector by attentional mechanism and passed into LSTM model and multi-layer perceptron for similarity prediction.In the comparison experiment of graph regression task and graph classifica-tion task,the model obtains eight optimal performance and six optimal performance respectively.
关键词
图神经网络/图相似性学习/可微池化/图匹配/相似性搜索/图编辑距离/注意力机制Key words
graph neural networks/graph similarity learning/DiffPool/graph matching/similarity search/graph edit distance/attention mechanism引用本文复制引用
基金项目
国家自然科学基金项目(61976032)
国家自然科学基金项目(62002039)
出版年
2024
NETL
NSTL
万方数据