南京大学学报(自然科学版)
南京大學學報(自然科學版)
남경대학학보(자연과학판)
JOURNAL OF NANJING UNIVERSITY(NATURAL SCIENCES)
2015年
1期
139-147
,共9页
多标记学习%正则化最小二乘分类%二分类问题%核函数%Sylvester方程
多標記學習%正則化最小二乘分類%二分類問題%覈函數%Sylvester方程
다표기학습%정칙화최소이승분류%이분류문제%핵함수%Sylvester방정
multi-label learning%regularized least squares classification%binary classification problem%kernel function%Sylvester equation
在传统的监督学习中,每个对象由单个实例表示且只属于一个类别标记。然而,在多标记学习中,每个对象由一个实例表示但可能同时属于多个类别标记,其任务是预测未知样本的类别标记集合。本文提出了基于正则化最小二乘的多标记分类算法,即将传统的正则化最小二乘分类推广到多标记学习中。首先,将多标记学习问题转化为多个独立的二分类问题(每个对应一个类别标记);其次,为了充分利用类别标记之间的相关信息,构建了基于类别标记的邻接图,其中每个节点代表一个类别标记,每条边的权重反映了相应类别标记对之间的相似性。最后,构建了建立在核函数基础上的多标记正则化最小二乘模型,并可以转化为求解一个 Sylvester方程。在8个基准数据集上用5种不同的评价准则进行度量的实验结果表明了本文算法优于其他6种常用的多标记分类算法。
在傳統的鑑督學習中,每箇對象由單箇實例錶示且隻屬于一箇類彆標記。然而,在多標記學習中,每箇對象由一箇實例錶示但可能同時屬于多箇類彆標記,其任務是預測未知樣本的類彆標記集閤。本文提齣瞭基于正則化最小二乘的多標記分類算法,即將傳統的正則化最小二乘分類推廣到多標記學習中。首先,將多標記學習問題轉化為多箇獨立的二分類問題(每箇對應一箇類彆標記);其次,為瞭充分利用類彆標記之間的相關信息,構建瞭基于類彆標記的鄰接圖,其中每箇節點代錶一箇類彆標記,每條邊的權重反映瞭相應類彆標記對之間的相似性。最後,構建瞭建立在覈函數基礎上的多標記正則化最小二乘模型,併可以轉化為求解一箇 Sylvester方程。在8箇基準數據集上用5種不同的評價準則進行度量的實驗結果錶明瞭本文算法優于其他6種常用的多標記分類算法。
재전통적감독학습중,매개대상유단개실례표시차지속우일개유별표기。연이,재다표기학습중,매개대상유일개실례표시단가능동시속우다개유별표기,기임무시예측미지양본적유별표기집합。본문제출료기우정칙화최소이승적다표기분류산법,즉장전통적정칙화최소이승분유추엄도다표기학습중。수선,장다표기학습문제전화위다개독립적이분류문제(매개대응일개유별표기);기차,위료충분이용유별표기지간적상관신식,구건료기우유별표기적린접도,기중매개절점대표일개유별표기,매조변적권중반영료상응유별표기대지간적상사성。최후,구건료건립재핵함수기출상적다표기정칙화최소이승모형,병가이전화위구해일개 Sylvester방정。재8개기준수거집상용5충불동적평개준칙진행도량적실험결과표명료본문산법우우기타6충상용적다표기분류산법。
In traditional supervised learning,each obj ect is represented by a single instance and associated with only one class label.Recently,multi-label learning attracts much attention from researchers of machine learning,pattern recognition etc.,due to its ability in obtaining more information on predicting multiple class labels.In multi-label learning,each obj ect is represented by an instance,while it may be assigned to multiple class labels simultaneously. Hence,its task is to predict a class label set for the unknown instance.The resulting model is usually an ill-conditioned quadratic program,which requires regularization.In this paper,a multi-label classification algorithm based on regularized least squares classification is proposed,which is derived from the traditional regularized least squares classification.To establish our model,we first transform the multi-label learning problem into multiple inde-pendent binary classification problems (each for one class label).Then,in order to fully exploit the class label correlations information,an adjacent graph based on all the class labels is built,where each node stands for one class label and the weight of each edge reflects the similarity between corresponding pairwise-label.Finally,a multi-label regularized least squares model is constructed on the basis of kernel function,whose first order optimality condition is a Sylvester equation,which can be solved efficiently by using numerical linear algebra technique.We perform experiment on eight benchmark data sets in terms of five different evaluation criteria and compare it with the other six state-of-the-art multi-label learning algorithms.The results show that our algorithm is competitive.