Some Elements of Machine Learning

Abstract

This talk will revisit some important elements of ML lore, focusing on the design of classifier-learning systems. Within ML, the key desiderata for such systems have been predictive accuracy and interpretability. Although Provost, Fawcett and Kohavi (1998) have shown that accuracy alone is a poor metric for comparing learning systems, it is still important in most real-world applications. The quest for intelligibility, stressed from earliest days by Michie, Michalski and others, is now crucial for those data-mining applications whose main objective is insight. Scalability is also vital if the learning system is to be capable of analyzing the burgeoning numbers of instances and attributes in commercial and scientific databases.

This extended abstract also appears in the Proceedings of the Sixteenth International Conference on Machine Learning, published by Morgan Kaufmann.