基于集合学习的超链接预测  

Hyperlink prediction based on collection learning

作　　者：刘臣[1] 袁慕婷 周立欣[1] LIU Chen;YUAN Muting;ZHOU Lixin(Business School,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学管理学院,上海200093

出　　处：《智能计算机与应用》2024年第4期162-167,共6页Intelligent Computer and Applications

基　　金：上海市哲学社会科学规划课题(2021BTQ003);中国博士后科学基金第69批面上项目(2021M692135)。

摘　　要：超图链接预测作为图预测的重要研究方向之一,能够通过预测节点间的高阶相互作用解决许多实际问题。目前大多数链接预测研究多集中于成对关联关系的预测,而实际应用中链接关系的对象往往大于两个。因此,本文提出一种基于集合表示和transformer的链接预测模型。该模型通过对集合表示实现链接预测的无序性,并将传统的语言模型拓展应用于链接预测问题。模型首先将数据嵌入编码层对数据特征进行提取,然后使用池化机制对解码层进行解码,并引入评分函数对模型预测结果进行评估。实验表明,本文提出的模型可以有效利用网络结构特征,在6个不同规模的代谢网络数据集上的表现优于多个基准算法。Hypergraph link prediction as one of the important research directions of graph prediction,it can solve many practical problems by predicting the high-order interaction between nodes.At present,most of link prediction studies focus on the prediction of pairwise association relations,but in practice,the objects of link relations are often more than two.Therefore,this paper proposes a link prediction model based on set representation and transformer,which realizes the disorder of link prediction through set representation,extends the traditional language model to link prediction problem.The model first embedded the data into the encoding layer to extract the data features,then used the pooling mechanism to decode the decoding layer,introduced the scoring function to evaluate the prediction results of the model.Experiments show that the model we proposed can effectively utilize the structural features of the network and outperform multiple benchmark algorithms on six metabolic network datasets of different sizes.

关 键 词：超链接预测 注意力机制 集合学习 

分 类 号：O157.5[理学—数学]