Research On Multiple-instance Boosting Algorithm For Self-paced Learning

Multiple-instance is one of the research hotspots in machine Learning.In order to establish a more accurate classifier,most of the multiple-instance learning algorithms require a large number of labeled multiple-instance bags.However,there is a huge economic cost to manually label a large number of bags.Therefore,it is an urgent problem to establish an accurate multiple-instance learning model when only a few bags are labeled,a large number of bags are not labeled,and only little labeling information is available(called "weak label multiple-instance learning problem").In recent years,researchers have proposed self-paced learning which solves the problem of missing data labels.Self-paced learning simulates a person's learning process.It first learns simple knowledge,and then learns complex knowledge.Its robust mechanism works well for dealing with label missing data.To solve the problem that only a few bags are labeled and a large number of bags are unlabeled,this paper incorporates self-paced learning and boosting into the multiple-instance learning process,and proposes a Self-Paced Boost Multiple Instance Learning(SP-B-MIL)algorithm.The main research work of this paper includes:1.This paper proposes a multiple-instance self-paced learning function,so that self-paced learning can be applied to solve multiple-instance learning problems.In each iteration,the self-paced learning function selects instances from different bags to ensure the diversity of instances.2.Based on the self-paced learning loss formula,this paper proposes a multiple-instance boosting model and designs an optimized solution for this model.Experimental results show that the algorithm proposed in this paper has better classification performance than the existing multiple-instance learning algorithms when there are only a few labeled bags in the training data set.