一种新颖的多实例学习算法与应用

A Novel Multi-instance Learning Algorithm and Its Application

多实例学习(MIL)作为一种半监督学习形式,其中训练数据标签上只有不完整的知识。具体而言,标签被分配在这些包上,包中实例的标签未知。在MIL算法中,如果包中至少有一个实例为正,则包被标记为正;如果包中的所有实例均为负,则包标记为负。MIL算法的目标是通过学习一个分类函数,预测测试数据中包或实例的标签。同时,MIL的性质使其可应用于多种应用,从药品活动预测到文本或多媒体信息检索。对多样化密度算法的缺陷进行了改进,提出了一种新颖的多实例学习算法。最后,在图像分类/检索问题数据集-Corel数据库上,将提出的算法与其他算法,进行了性能对比评估。

As a form of semi-supervised learning in which training data Multi-instance learning,MIL has only incomplete knowledge on the label.Specifically,labels are assigned to these packages,and the labels of the instances in the package are unknown.Currently,MIL has become an active research area.In the MIL algorithm if at least one instance in the bag is positive,the bag is marked as positive;if all instances in the bag are negative,the bag is marked as negative.The goal of the MIL algorithm is to predict the label of a bag or instance in the test data by learning a classification function.At the same time,the nature of MIL makes it applicable to a variety of applications,from drug activity prediction to text or multimedia information retrieval.In this paper,the shortcomings of the diversity density algorithm were improved,and a novel multi-instance learning algorithm was proposed.Finally,on the image classification/retrieval problem dataset-Corel Database,the performance of the proposed algorithm was compared with other algorithms.