| In the era of big data,the cost of labeling data has increased significantly.The multiinstance learning paradigm can greatly reduce the requirements for data labels,so it has become a current research hotspot.This paradigm names the collection of multiple examples as ”packages” and gives all samples in the package a common label.In the era of big data,the cost of labeling data has increased significantly.The multi-instance learning paradigm can greatly reduce the requirements for data labels,so it has become a current research hotspot.This paradigm names the collection of multiple examples as ”packages”and gives all samples in the package a common label.In the era of big data,the cost of labeling data has increased significantly.Multi-instance learning paradigm can greatly reduce the requirements for data labels,which has become a current research hotspot.This paradigm names the collection of multiple instances as ”bags” and gives all instances in the bag a common label.On the premise of a large amount of complete training data,existing multi-instance learning researches have achieved good results.However,in practical problems,due to the labeling cost and environmental constraints,the ideal labeling data set is often difficult to obtain.Compared with data with complete labels,the labeling cost required to collect unlabeled data or unilateral positive or negative data is greatly reduced.Therefore,the multi-instance learning problem in the unilateral labeling scenario can further reduce the requirements for data labeling,and has a wide range of applications,which is of practical significance.This is a novel learning problem that has received little attention at present.This paper studies the unilateral labeling problem in multi-instance scenarios from two perspectives,and proposes different solutions respectively.The main content and work of the article are as follows:(1)Multi-instance learning method on negative-unlabeled data based on two kinds of information mapping In the multi-instance learning scenario with only negative labels,the multi-instance problem is transformed into a single-instance problem through mapping and then solved.In the mapping step,in order to make full use of the information of the one-sided label data,the statistical features based on the data information and the discriminative mapping features based on the label information are considered at the same time.First,use statistical features to assign a ”pseudo-label” to each unlabeled bag,so as to further extract the key information of each instance and obtain more distinguishable mapping features.After ensuring that the data information is fully extracted in the mapping step,the empirical risk minimization loss function is used to obtain the optimal classifier that is unbiased in the overall distribution of the data from only the one-sided labeled training data.Finally,experiments on multiple data sets verify the effectiveness of our method.(2)Multi-instance learning method on positive-unlabeled data based on bi-level embedding Compared with the one-sided negative label data,the multi-instance learning problem with only positive labels is more difficult.In the multi-instance scenario,not all the instances in the positive bag are positive instances,which results in the information contained in the only positive label being more ambiguous.Aiming at the problem of multi-instance learning in this scenario,this paper proposes a multi-instance learning method based on bi-level embedding.This method solves the problem by transforming the multi-instance problem into a single-instance problem.Different from the simple singlelayer mapping method,the bi-level embedding mechanism develops different mappings for positive bags and unlabeled bags,and adopts a weighting strategy in the positive instances to ensure that the interference of negative instances can be eliminated,so that the characteristics of the key examples in positive bag are highlighted.The bi-level embedding strategy can extract the information of the purified key instances,and remove the ambiguity of the label information to a large extent.Finally,on the basis of single samples,the optimal classifier is obtained by using the classification error minimization loss function based on the class prior probability.Through experiments on multiple data sets,the effectiveness of our method is verified. |
