The Metric Learning Algorithms Based On Universum Learing

In recent years,with the rapid development of computer technology,the Internet has become an indispensable part in people's daily life.The rapid development of the Internet brings convenience to people's lives,and it also brings big data with infinite value.Machine learning is a discipline that exploits the value of big data.Machine learning algorithms have a large number of applications in real life,such as the face recognition unlocking in mobile phone,Speech Recognition,automatic translation and auto driving,etc.People's lives are increasingly inseparable from the convenience of machine learning algorithms.These machine learning algorithms such as SVM,KNN,and K-Means rely on the distance measurement between samples.Euclidean distance is the most common distance metric,but the distance metric like Euclidean distance only dependent on data and task-independent is not always effective.How to comprehensively utilize the given data and specific tasks to learn a new distance metric to better accomplish a given task is precisely the research goal of metric learning.Many researchers have already studied this.Geometric Mean Metric Learning(GMML)is an efficient metric learning algorithm.GMML learns a metric matrix A so that the distance between similar sample points is as small as possible and the distance between them is as large as possible under the metric matrix A.The training samples used by GMML are all target data,but there are a large number of non-target data in the same field which called Universum data is not utilized,it inevitably leads to waste of information.In fact,Universum data can provide a certain amount of prior knowledge in the field.Based on this,we proposed a new metric learning algorithm based on Universum learning,which called U-GMML.U-GMML expects to obtain a new metric matrix A,so that the distance between similar samples is as small as possible,the distance between samples of different classes is as large as possible,and the distance between Universum data and target class data is as large as possible,thus Make the learned metric matrix A more favorable for classification.This paper further combines metric learning with multi-view learning and further generalizes U-GMML to multi-view data.A multi-view metric learning algorithm which called MU-GMML based on Universum learning is proposed.MU-GMML hopes to learn a metric matrix A.Under this metric matrix A,the distance between similar samples of different views is as small as possible,and the distance between samples of different classes of different views is as large as possible.At the same time,the distance between the target data of any view and the Universum data of another view is as large as possible.The experimental results on the real data set verify the effectiveness of the two algorithms.