Learning Instance Weights in Multi-Instance Learning

Abstract

Multi-instance (MI) learning is a variant of supervised machine learning, where each learning example contains a bag of instances instead of just a single feature vector. MI learning has applications in areas such as drug activity prediction, fruit disease management and image classification.

This thesis investigates the case where each instance has a weight value determining the level of influence that it has on its bag's class label. This is a more general assumption than most existing approaches use, and thus is more widely applicable. The challenge is to accurately estimate these weights in order to make predictions at the bag level.

An existing approach known as MILES is retroactively identified as an algorithm that uses instance weights for MI learning, and is evaluated using a variety of base learners on benchmark problems. New algorithms for learning instance weights for MI learning are also proposed and rigorously evaluated on both artificial and real-world datasets. The new algorithms are shown to achieve better root mean squared error rates than existing approaches on artificial data generated according to the algorithms' underlying assumptions. Experimental results also demonstrate that the new algorithms are competitive with existing approaches on real-world problems.