多粒度自适应嵌入融合的有向超图表示学习模型  

Directed Hypergraph Representation Learning with Multi-granularity Adaptive Embedding Fusion

作　　者：马紫彤 赵文博 杨哲[2] MA Zitong;ZHAO Wenbo;YANG Zhe(School of Computer Science and Technology,Soochow University,Suzhou 215006,China;Provincial Key Laboratory for Computer Information Processing Technology,Suzhou 215006,China)

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006 [2]江苏省计算机信息处理技术重点实验室,江苏苏州215006

出　　处：《小型微型计算机系统》2025年第3期586-593,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61772356,61876117)资助;江苏高校优势学科建设工程项目资助.

摘　　要：图表示学习能够挖掘图结构数据中蕴含的丰富信息,例如结构、关系信息等.近年,涌现了大量针对高阶、复杂图结构的表示学习研究,然而针对高阶有向超图结构的研究相对有限,且存在一定的局限性:现有方法无法同时提取有向超图的高阶性和方向性,导致其失去了结构优势.同时,在图表示学习中,信息通过连接边实现信息传播,堆叠网络层数时容易产生过平滑问题.为解决上述问题,本文首先设计有效且能够在通用的有向超图结构中提取信息的卷积模块,在避免信息损失下有效地传递结构信息;其次采用自适应权重的嵌入融合机制,来缓解过平滑问题.在多个不同类型的数据集上的实验表明了有向超图表示学习模型的先进性,在分类任务上的准确率最高提升4.39%.Graph representation learning has proven adept at extracting nuanced information from graph-structured data,encompassing crucial structural and relational intricacies.In recent years,a prolific emergence of research endeavors has been noticed,particularly focusing on high-order and intricate graph structures.However,there has been limited scholarly inquiry into high-order directed hypergraph structures,coupled with certain inherent limitations:directed hypergraphs simultaneously possess the high-order attributes characteristic of hypergraphs and the directional properties reminiscent of directed graphs.However,owing to their intricate structural intricacies,the utilization of existing representation learning methodologies often entails substantial information loss,culminating in the attenuation of their structural distinctiveness.Additionally,in graph representation learning,information is propagated through connecting edges.However,when stacking multiple network layers,the issue of over-smoothing commonly arises,impeding the model′s aptitude for uncovering intricate spatial relationships.To address the aforementioned challenges,we first propose an effective directed hypergraph convolution module designed to efficiently convey structural information while avoiding information loss.Secondly,we employ an embedding fusion mechanism with adaptive weights to mitigate the over-smoothing problem.Experiments conducted on multiple diverse datasets have demonstrated the superiority of the directed hypergraph representation learning model.The model achieved its highest accuracy improvement of 4.39%in node classification tasks.

关 键 词：有向超图 表示学习 有向超图卷积 自适应嵌入融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]