Machine Learning and Friends Lunch





home
past talks
resources
conferences

Untangling Multiple-Instance Learning


Amy McGovern
UMass

Abstract


In this talk, I will present an 'untangled' view of multiple-instance learning (MIL) which separates representation, search, and model evaluation. We motivate the use of the MIL framework with three real-world example tasks based on classifying robot grasps, identifying musky molecules, and discovering predictive structures in relational data. Untangling MIL allows us to clearly compare existing approaches as well as enabling the use of traditional supervised learning techniques to solve MIL tasks. In particular, we show that the chi-squared statistic can be used as an evaluation function and that it provides the ability to prune the search in a guaranteed manner. The use of a ranking classifier enables us to make informed choices about classification thresholds using receiver-operator curve (ROC) analysis. We show the decomposition of well-known approaches such as diverse density (Maron & Lozano-Perez, 1998 and Maron, 1998) and axis-parallel rectangles (Dietterich et al., 1997). We present experimental results using a standard two-dimensional data set, the MUSK task, and the Internet Movie Database.

Back to ML Lunch home