Research On Metric Learning Algorithm Based On Sample Pairs And Triples

With the advent of the information age and the gradual improvement of network technology,machine learning,as an important branch of scientific computing,has made great progress in the field of artificial intelligence.Among them,metric learning has become a hot research field in machine learning because it can better obtain the similarity relationship between samples.Metric learning can adaptively learn metrics from samples to complete specific tasks.It can not only improve the performance of algorithms,such as k-means,k-nearest neighbor,SVM,and so on,but also be widely used in many fields,such as image retrieval,target detection,face recognition,and so on.Therefore,metric learning has certain theoretical value and research significance.In this paper,the topic of "Research on metric learning algorithm based on sample pairs and triples" is studied,and the research background,research status,and classical metric learning algorithms of metric learning are described.Because of the problem of what kind of loss function is constructed by Mahalanobis distance metric learning and how to construct effective constraints to improve algorithm performance,the following research is conducted:(1)The metric learning based on binary constraints mostly expands the feasible region by introducing slack variables.Metric learning algorithms generally establish relaxation variables for each constraint from a micro perspective.When the value of relaxation variables is large,the model will deviate,resulting in poor generalization performance of the algorithm.To solve this problem,this paper proposes a metric learning algorithm based on global relaxed variables.The algorithm establishes relaxation variables for similar sample sets and dissimilar sample sets from macro perspective.The proposed algorithm is compared with the representative metric learning algorithm.The experimental results show that the proposed algorithm can improve the classification performance to a certain extent by considering the relaxation variables from a macro perspective.(2)Metric learning based on triple constraint is an important metric learning algorithm.The triple constraint not only considers the similarity relationship between samples within a class but also considers samples between classes.Most triples are constructed based on prior knowledge,which has a certain influence on algorithm performance.To solve this problem,this paper proposes a metric learning algorithm with adversarial sample triples.Using the idea of adversarial training for reference,the algorithm learns adversarial samples near the original samples to construct adversarial triples constraints for training.The experimental results show that the proposed algorithm not only overcomes the problem that the existing fixed constraint methods are greatly affected by prior knowledge,indicates that the algorithm can improve the performance by distinguishing the more difficult triples to a certain extent.In this paper,based on Mahalanobis distance,a metric learning algorithm based on global relaxation variables and a metric learning algorithm with adversarial sample triples constraints are presented respectively.The experimental results show the superiority of the proposed algorithm.