基于SCL-CMM模型的超网络链路预测方法

Hypernetwork link prediction method based on the SCL-CMM model

图为建模现实系统的内在交互提供了一种有效的方式,但却无法显示捕获的多个实体之间广泛存在的高阶异质性,超图则可以很好地突破低阶关系的限制。超网络的链路预测就是根据观测到的超图结构来预测未知的超链路,因其可以充分地刻画复杂系统的关联模式而成为网络科学中的热点问题。现有的方法通常针对整个拓扑结构设计推理模型,忽略了网络中隐含的聚集特性,导致预测的超链路类别不全面。针对上述问题,提出了基于超图谱聚类解析器的协调矩阵最小化(coordination matrix minimization based on hyper graph spectral clustering parser,SCL-CMM)模型的超网络链路预测方法。该方法将高阶超网络映射到具有一定语义的异质超图上,然后利用谱聚类解析器来提取超链路的结构特征,将原始超图重构为多个同质子图,进而在子图的观测空间而不是整个网络的邻接空间推断潜在超链路的分布情况,还原完整的超网络结构。该方法联合学习超网络的结构特征与集聚属性来建模各个子图的高阶非线性行为,解决了异构超图链路预测类别单一、精度低的问题。在9个真实数据集上进行了大量的对比实验表明,该方法在AUC(area under curve)评分和召回率方面都显著优于现有方法。

An effective method for the internal interaction within modeling reality systems is provided by graphs;however,they have been unable to effectively display and capture the high-order heterogeneity that widely exists between multiple entities.Hypergraphs have been recognized for their ability to surpass the limitations imposed by low-order relationships.Hypernetwork link prediction,which involves predicting unknown hyperlinks based on the observed hypergraph structure,has increasingly become a hot topic in network science due to its capacity to fully describe the association patterns of complex systems.Existing methods typically design reasoning models for the entire topology,often overlooking the implicit aggregation characteristics within the network,which leads to an in‐complete prediction of hyperlink categories.To address these issues,a coordination matrix minimization model based on hypergraph spectral clustering parser(SCL-CMM)was proposed.Initially,higher-order hypernetworks were mapped into heterogeneous hypergraphs with certain semantics.Subsequently,the spectral clustering parser was employed to extract the structural features of hyperlinks.The original hypergraph was reconstructed into mul‐tiple homoprotonic graphs,and the distribution of potential hyperlinks was inferred within the observation space of the subgraph,rather than the entire adjacency space,in order to restore the complete hypernetwork structure.This method federated learned the structural characteristics and aggregation attributes of hypernetworks to model the high-order nonlinear behavior of each subgraph,thereby solving the problems of single category and low precision in het‐erogeneous hypergraphs link prediction.Extensive comparative experiments were conducted on nine real datasets,demonstrating that this method significantly outperformed existing methods in terms ofAUC score and recall rate.