一种基于SCAD的改进谓词发现方法  

An improved predicate invention method based on SCAD

作　　者：郑晓东[1] 潘敬敏 胡汉辉[1] ZHENG Xiaodong;PAN Jingmin;HU Hanhui(School of Economics and Management,Southeast University,Nanjing,Jiangsu 211189,China;Southeast University-Monash University Joint Graduate School(Suzhou),Suzhou,Jiangsu 215000,China)

机构地区：[1]东南大学经济管理学院,江苏南京211189 [2]东南大学-蒙纳士大学苏州联合研究生院,江苏苏州215000

出　　处：《江苏大学学报（自然科学版）》2018年第5期576-580,610,共6页Journal of Jiangsu University：Natural Science Edition

基　　金：国家自然科学基金资助面上项目(70673010)

摘　　要：针对归纳逻辑编程中传统谓词发现方法会导致错误级联的难题,提出一种基于平滑削边绝对偏离(smoothly clipped absolute deviation,SCAD)正则化稀疏的改进谓词发现方法.新方法并不明确地创建新谓词,而是通过使用正则化稀疏方式将软谓词发现的参数一起正则化,从而隐式地组合紧密相关的规则.在软谓词发现中引入SCAD这一正则化稀疏模型,同时针对无偏稀疏性,着重观察SCAD对软谓词发现结果的影响.基于欧洲皇室家庭关系数据集进行试验,确定了μ,α的最优值,并完成了知识库完善试验.结果表明,该方法能有效克服错误级联这一难题,缩短对知识库的查询时间,并可提高谓词发现的平均精准度到0.798,远超过基于拉普拉斯正则化的软谓词发现方法的0.726.To solve the problem of error cascades of traditional predicate invention method in inductive logic programming(ILP),an improved approach was proposed based on smoothly clipped absolute deviation penalty(SCAD)regularized sparsity.Instead of explicitly creating new predicates,the predicate invention method implicitly grouped closely-related rules by regularized sparsity to regularize the parameters together.The regularized sparse model of SCAD was introduced into the soft predicate invention.For the unbiased sparseness,the influence of SCAD on soft predicate invention results was analyzed.The experiments were completed based on the dataset of European royalty family ties,and the values ofμandαwere determined to improve the knowledge base.The results show that the proposed approach can effectively overcome the difficulty of error cascades,and the method can improve the mean average precision in predicate invention and shorten the query time of knowledge base.The mean average precision of the approach based on SCAD regularized sparsity is 0.798,which is more higher than the soft predicate invention method of 0.726 based on Laplacian regularization.

关 键 词：谓词发现 正则化稀疏 SCAD 归纳逻辑编程 假设语言 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]