Research On Multi-Instance Learning Methods With Sample Distribution Change

Multi-instance learning is a variant of supervised learning framework which has received a considerable amount of research attention due to its power in describing complicated learning concepts. With many algorithms and theoretical results developed in the past decades, multi-instance learning has been applied to real-world tasks such as image classification, video annotation, web page categorization, drug activity prediction, etc. Existing multi-instance researches assume that training and testing samples are independently drawn from the same distribution. However, training and testing distributions may be different in real-world tasks. How to solve multi-instance learning with distribution change is still an open question.This thesis studies multi-instance learning with distribution change, and achieves the following main innovative contributions:1. We propose a new algorithm MICS, which can solve multi-instance learning with distribution change prolem effectively. By utilizing the relationship among instances in the same bag, MICS corrects covariate shift at both instance-level and bag-level, experiments verifies the effectiveness of MICS.2. We analyzed the influence of instance relationship on the learnability of multi-instance learning, the result shows that the sample complexity of multi-instance learning is logarithmically dependent on the maximum bag size.