Industrial and Systems Engineering
University of Florida

EIN 6918: Graduate Seminar
Spring 2008

April 17, 2008

3PM, MAEB 211

Multiple Instance Learning via Margin Maximization

Erhun Kundakcioglu

Department of Industrial and Systems Engineering

University of Florida

Abstract

Machine learning has been one of the fastest growing subfields in theoretical computer science with many applications ranging from text categorization to failure prediction. The exponential growth in the size of databases requires reliable automatic detection and prediction algorithms, which are efficient and fast. There are numerous support vector machine (SVM) methods and implementations that carry out such tasks with great success. Multiple instance learning (MIL) refers to classification of unseen bags based on the labeled bags of instances as the training data. Despite the large number of SVM based classification methods, there are only a few SVM based MIL methods in the literature. We first consider the margin maximization problem for multiple instance data and prove that the problem is NP-hard. We also propose an iterative SVM based MIL method and present computational results on publicly available databases. Our approach outperforms three previous SVM based MIL methods in a majority of image annotation, text-categorization, and drug activity prediction data sets.

[Joint work with Onur Seref and Panos M. Pardalos]