基于可微池化的层级图相似性学习  

Hierarchical graph similarity learning based on differentiable pooling

作　　者：吴磊 李晓楠 李冠宇[1] WU Lei;LI Xiao-nan;LI Guan-yu(College of Information Science and Technology,Dalian Maritime University,Dalian 116026,China)

机构地区：[1]大连海事大学信息科学技术学院,辽宁大连116026

出　　处：《计算机工程与设计》2024年第7期2013-2020,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(61976032、62002039)。

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。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 classification 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 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]