Decision Trees

In the early 1990s a surgeon and two computer scientists at the University of Wisconsin, William Wolberg, Nick Street and Olvi Mangasarian, digitized microscope images of fine-needle aspirates: small samples of cells drawn from breast masses. For each of 569 masses they measured ten properties of the cell nuclei, such as radius, texture and the number of concave points on the outline, and recorded each as a mean, a standard error and a worst value, 30 numbers per patient. Their classifier was a decision tree of an unusual kind, built by linear programming, whose single split was a tilted plane through three of the measurements. The dataset's documentation, dated November 1995, reports an estimated accuracy of 97.5% and adds that the classifier "has correctly diagnosed 176 consecutive new patients".