超图学习综述:算法分类与应用分析

Survey on Hypergraph Learning:Algorithm Classification and Application Analysis

随着图结构化数据挖掘的兴起,超图作为一种特殊的图结构化数据,在社交网络分析、图像处理、生物反应解析等领域受到广泛关注.研究者通过解析超图中的拓扑结构与节点属性等信息,能够有效解决实际应用场景中所遇到的如兴趣推荐、社群划分等问题.根据超图学习算法的设计特点,将其划分为谱分析方法和神经网络方法,根据方法对超图处理的不同手段,可进一步划分为展开式方法和非展开式方法.若将展开式方法用于不可分解超图,则很有可能会造成信息损失.然而,现有的超图相关综述文章鲜有就超图学习方法适用于哪类超图这一问题做出相关归纳.因此,分别从超图上的谱分析方法和神经网络方法两方面出发,对展开式方法和非展开式方法展开讨论,并结合其算法特性和应用场景作进一步细分;然后,分析比较各类算法的设计思路,结合实验结果总结各类算法的优缺点;最后,对超图学习未来可能的研究方向进行了展望.

With the rise of graph structured data mining,hypergraph,as a special type of graph structured data,is widely concerned in social network analysis,image processing,biological response analysis,and other fields.By analyzing the topological structure and node attributes of hypergraph,many problemscan be effectively solved such as recommendation,community detection,and so on.According to the characteristics of hypergraph learning algorithm,it can be divided into spectral analysis method,neural network method,and other method.According to the methods used to process hypergraphs,it can be further divided into expansion method and non-expansion method.If the expansion method is applied to the indecomposable hypergraph,it is likely to cause information loss.However,the existing hypergraph reviews do not discuss that hypergraph learning methods are applicable to which type of hypergraphs.So,this article discusses the expansion method and non-expansion method respectively from the aspects of spectral analysis method and neural network method,and further subdivides them according to their algorithm characteristics and application scenarios.Then,the ideas of different algorithms are analyzed and comparedin experiments.The advantages and disadvantages of different algorithms are concluded.Finally,some promising research directionsare proposed.