Research On Multi-Instance Multi-Label Algorithm Based On SVM And Privileged Information

In a natural scene image,a part of the image can be represented as an instance,such as the sun,mountains,or trees.Therefore,the classification of the natural scene image involves a multiple instances multiple labels learning framework.In the text classification task,articles may belong to multiple categories,and can be represented by a bag of instances,and an instance represents a paragraph.Multiple instances and multiple labels learning is a new framework for such complex objects.In the past decade,many methods have been proposed to solve the problem of multiple examples and multiple labels learning.But most of them simply adopt the strategy of reducing the complexity of the problem,which is to transform these multiple instance multiple label learning(MIML)problems into traditional single instance single label learning(SISL)problems.In the process of transformation,the introduction of privilege information is not considered,which makes it difficult to guarantee the full and reasonable description of the relationship between instances and the relationship between instances and tags.In the multiple instances and multiple labels learning problem studied in this paper,this method studies whether the algorithm has more advantages after giving information about other aspects of the training data set,which is called learning to use privileged information.In this paper,we mainly study the multiple instance multiple label learning algorithm based on the additional privilege information of SVM,and propose a new method called SMP,which uses the privilege information(PI)for multiple instance multiple label.The main research contents are as follows(1)How to find the most "suitable" single label for each example.This method(SMP)makes full use of the relationship among examples,privilege information(PI)and tags in multiple instance and multiple tag instances to find the most suitable single label belonging to this group of labels for all examples in the multiple instances multiple labels instance package(2)How to combine the privilege information to help us find the best interval hyper plane.Privilege information can be incorporated into multiple instance multiple label learning(MIML).In the first step,the quadratic programming is proposed,and the empirical risk function is minimized by L2 norm regularization.In the second step,we use privilege information to assign a single label to each example to implement integer programming.In order to solve the algorithm proposed in this paper,we use alternating optimization method and random gradient algorithm(3)In order to verify the effectiveness of the method proposed in this paper,we have carried out experiments on scene data,Microsoft Research Cambridge(MSRC)image data and Pascal visual object CALS challenges(VOC)2012 subset.The experimental results show that this method has more advantages than other multiple example multiple label algorithms.