Abstract

Intelligent user interfaces increasingly employ machine-learning components in order to learn about individual users and optimize the interface accordingly. In all of these cases, the learning tasks must be defined before deployment and a machinelearning expert must be involved in the development process. Unfortunately, this methodology significantly limits the potential utility of machinelearning in intelligent user interfaces since there is no way for a user to create new learning components that serve specific needs as they arise. Our work addresses this shortcoming by developing a framework for user-initiated learning (UIL), where the end user is able to define new learning tasks in a natural way, after which the system automatically generates a learning component, without the intervention of a machine-learning expert. In particular, we consider a UIL scenario where the user is able to define learning problems for predicting when certain activities are forgotten by the user (eg encrypting an email message), upon which the system can generate warnings. We describe the knowledge representation and reasoning mechanisms required to replace the machine-learning expert, so as to automatically generate training examples, label the examples, select legal and relevant features, and finally learn the required concept. We present an implementation of this approach in the context of a popular email client and give initial experimental results.

Date

2009

Authors

Kshitij Judah, Thomas Dietterich, Alan Fern, Jed Irvine, Michael Slater, Prasad Tadepalli, Melinda Gervasio, Christopher Ellwood, William Jarrold, Oliver Brdiczka, Jim Blythe

Journal

IJCAI Workshop on Intelligence and Interaction, Pasadena, CA, USA