广义多线性混合效应模型  

Generalized multilinear mixed-effects model

作　　者：李超[1] 郭黎利[1] 窦峥[1] LI Chao;GUO Lili;DOU Zheng(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001

出　　处：《哈尔滨工程大学学报》2018年第5期934-940,共7页Journal of Harbin Engineering University

基　　金：国家自然科学基金项目(61671167)

摘　　要：为了解决多数据集间联合特征提取时数据分布多样、集间相关性结构复杂和共享特征方法多样的问题,本文提出了广义多线性混合效应模型。作为一种非监督多数据集特征提取方法,本算法可挖掘多个数据集之间的共享信息,实现对多数据集全局、局部和个体特征的提取。本算法利用了传统的广义线性模型,使其可以处理不同分布的张量数据集;并提出了一种基于超图的关系模型。该模型利用关系矩阵可以实现对数据集间相关结构的建模;通过提出辅助模式的概念,实现了特征的自动归类。数值实验结果表明:利用本算法提取的特征不仅反映了多数据集间的共同与个体信息,并且在人脸识别和推荐系统等问题中性能优于传统算法。To overcome difficulties in extracting the joint features among multiple datasets,including diversified data distribution,complex dependency structures among datasets,and diversified share feature methods,this paper proposes a generalized multilinear mixed-effects model as a novel method for extracting the features of nonsupervised multiple datasets. The algorithm can explore shared information among multiple datasets and extract global,partial,and individual features of multiple datasets. First,the algorithm popularized the traditional generalized linear model and enabled the processing of the tensor dataset with different distributions. Next,the algorithm proposes a relation model based on a hypergraph. By utilizing the relation matrix,modeling relevant structures among various datasets can be realized. Finally,the paper proposes the concept of an auxiliary mode and consequently realizes the automatic classification of features. Results of the numerical analysis demonstrate that the features extracted by utilizing the algorithm reflect the common and individual information within multiple datasets. In addition,the algorithm outperforms other traditional algorithms in the aspects of face recognition and system recommendation.

关 键 词：广义线性模型 张量分解 特征提取 多数据集学习 超图模型 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]