LHP: Logical hypergraph link prediction
LHP: Logical hypergraph link prediction
复制标题
DOI：
10.1016/j.eswa.2023.119842
复制
发表时间：
2023-03
期刊：
影响因子：
--
通讯作者：
Yang Yang-Yang;Xue Li;Yi Guan;Haotian Wang;Chaoran Kong;Jingchi Jiang
中科院分区：
文献类型：
--
作者：
Yang Yang-Yang;Xue Li;Yi Guan;Haotian Wang;Chaoran Kong;Jingchi Jiang
逻辑知识挖掘已经变得越来越重要，因为它是精确推理的基础，例如临床诊断或化学反应发现。知识超图为表达超越两两关联的固有高阶关系提供了一种自然格式。虽然逻辑知识（一阶逻辑中的合算符）显然是一种可以用超图表示的高阶关系，但以这种形式表示逻辑知识并通过超链接预测完成知识的方法尚未探索。在本研究中，逻辑知识由有向超边表示，并通过神经网络中的几何运算有效量化。提出的逻辑超链接预测器（LHP）利用逻辑“连接”操作的排列不变量和信息聚合等逻辑知识特征。LHP将逻辑表示捕获为一个整体单元，其中包含嵌入在单独超边中的不同关系类型，使其成为集成包含在超边中的逻辑知识的新方法。我们在他们的iaf1260b、iJO1366、USPTO和我们的中国医学高阶关系（CMHR）数据集上进行了广泛的实验。与最先进的超链接预测方法相比，LHP在线性复杂度方面取得了最好的性能，在接收器工作特性曲线（AUC）下的平均面积，macro_f1和精度方面。它在CMHR上特别有效，证明了逻辑知识表示在医学领域的重要性。LHP在CMHR上执行超边缘关系分类的平均AUC也达到了0.924。
Logical knowledge mining has become increasingly important as it is the foundation of precise reasoning, such as for clinical diagnosis or chemical reaction discovery. Knowledge hypergraphs provide a natural format for expressing inherently high-order relationships beyond pairwise associations. Although logical knowledge (the conjunct operator in first-order logic) is apparently a higher-order relation that can be represented by a hypergraph, methods of representing logical knowledge in this form and completing knowledge by hyperlink prediction have not yet been explored. In this study, logical knowledge is represented by directed hyperedges and effectively quantified by geometric operations in a neural network. The proposed logical hyperlink predictor (LHP) leverages logical knowledge features including permutation invariant and information aggregation of the logical ‘conjunct’ operation. LHP captures logical representation as a whole unit containing the different relationship types embedded within separate hyperedges, making it the novel method to integrate logical knowledge contained within hyperedges. We conduct extensive experiments on theiAF1260b,iJO1366, USPTO,and our Chinese Medical High-order Relational (CMHR) dataset. LHP achieved best performance with linear complexity compared to the state-of-the-art hyperlink prediction methods on a mean area under the receiver operating characteristic curve (AUC), Macro_F1and accuracy. It was particularly effective on the CMHR demonstrating the importance of logical knowledge representation in the medical field. LHP also achieved a mean AUC of 0.924 in performing hyperedge relationship classification on the CMHR.