Working Group on Theory and Method

Jim Berger, Patrick Wolfe, Jiayang Sun, Merlise Clyde, Andrew Noble, and others reraised an issue implicit in several of the presentations: many data mining methods that work well use much more than a minimal orthogonal basis of functions in doing their fits and predictions. Traditional statisticians find this unsettling---it seems to create a need for regularization or shrinkage, generates multiple testing problems, and can prevent the discovery of interpretable structures. The group would like to examine the issues to find out why gross expansion of the set of fitting functions seems to work. Starting Point: Susie Bayarri is making phone calls and a google search to prepare a short summary of the state of current knowledge. There's no deadline, but I expect she will be done in about two weeks.