Research On Multiple-instance Transfer Learning Method With Weak Labels

Multiple-instance learning is a new learning framework that has attracted more and more attention in the field of machine learning.In multiple-instance learning,the training set consists of a set of multiple-instance bags,and a multiple-instance bag contains several instances.If there is at least one positive instance in the multiple-instance bag,the multipleinstance bag is a positive bag,otherwise,the multiple-instance bag is a negative bag.Multiple-instance learning faces the following two challenges: 1)when labeling multipleinstance bags,different experts may give different classification labels to the same multipleinstance bag.How to construct a multiple-instance learning classifier when a multiple-instance bag corresponds to multiple inconsistent and non-uniform classification labels(called "weak labels"),and the actual classification label is unknown;2)when the number of labeled multiple-instance bags is relatively small,how to use related classification tasks to improve the model accuracy of the target task.For the challenges faced by multiple-instance learning,this paper proposes a multipleinstance transfer learning model based on weak labels.On the one hand,for the problem of weak labels in multiple-instance learning,we give each weak label a weight,and use the weighted weak labels to represent the labels of multiple-instance bags,and establish a multiple-instance learning classification model for weak labels.On the other hand,for the problem that the number of multiple-instance bags of the target task is relatively small,we introduce the task(“source task”)related to the target task into the learning process.On the basis of above-mentioned weak labels classification model,we merge information of the target task and the source task,and establish a multiple-instance transfer learning model based on weak labels to achieve knowledge transfer from the source task to the target task.In order to solve the model,firstly,the weights of weak labels are initialized in a random or equal way,and the weighted weak labels are used to represent the labels of multiple-instance bags.Secondly,based on the labels of multiple-instance bags,a multiple-instance transfer learning classifier is solved.Again,based on the result of the classifier,the weights of weak labels are updated.Finally,the weights of weak labels and classifier are updated through interactive iteration until the algorithm converges,and the final classification model is obtained.This paper conducts experiments on five actual data sets,and compares with the existing multiple-instance learning methods to verify the effectiveness of the multiple-instance transfer learning method based on weak labels.Experimental results show that the proposed method is superior to the existing multiple-instance learning methods in terms of classification accuracy and AUC(Area Under the Curve).