Seminar on Computational Learning and Adaptation




Tubular Neighbors for Regression and Classification


Art B. Owen
Statistics Department
Stanford University
art@playfair.stanford.edu



The simple k-nearest neighbor (k-NN) method is often very effective, especially in classification methods. When the number d of predictors is large, the nearest neighbors are likely to be quite distant from the target point. Furthermore they tend to all be on one side of the target point. Thus, while k-NN sounds like interpolation, it very often is an extrapolation. Worse still, this extrapolation can easily be in a direction for which the neighbor set is an unusually poor experimental design. This work introduces an axial-radial decomposition of the space around the target point. Prediction at the target involves interpolation in d-1 (radial) directions and extrapolation in at most 1 (axial) direction. The interpolation and extrapolation directions are treated differently when designing a neighborhood and when selecting a model to fit over the neighborhood. The neighborhood shapes are chosen from a one parameter family of ellipsoidal tubes, by cross-validation, along with the neighborhood size k. The resulting method is competitive on some classification problems from the U.C. Irvine data repository.


Date: Thurs., April 2; Time: 4:15-5:30PM; Place: Gates 100


The goal of this seminar is to increase communication among local researchers with interests in computational approaches to learning and adaptation. If you would like to be added to (or removed from) the mailing list, or if you are interested in giving a talk in the seminar, please send email to iba@isle.org.


Return to seminar schedule.