A Case Study in Machine Learning

Empirical machine learning develops methods for formulating predictive theories from observations, drawing on ideas from arti cial intelligence and statistics. This paper discusses examples of such methods in the context of modelling the biomedical activity of a family of compounds. Five approaches are compared: standard multivariate linear regression, piecewise linear models expressed as model trees, a simple instance-based predictor, a combination of instance-based learning with model trees, and a hybrid of instance-based and analogical learning. The methods are evaluated on the predictive accuracy of the theories they construct for this task.