Seminar on Computational Learning and Adaptation



 
The Matching Representation for Tree Structures 
and its Applications to Tree-Based Classifiers

Susan Holmes
Department of Statistics
Stanford University
Stanford, CA 94305-4065
susan@stat.stanford.edu


A challenging task for computational statisticians involves the construction of taxinomic hierachies or trees from unsupervised training data. However, previous work in this topic typically lacks a well-defined measure of the distance between alternative trees or a means for searching the space effectively. This talk presents a natural coordinate system for semi-labeled binary trees using a correspondence with the set of perfect matchings in the complete graph.  This correspondence produces a distance between trees, and a way of enumerating all trees in a minimal step order.  It is useful in randomized algorithms as it enables moves on the space of trees that make random optimization strategies `mix' quickly.  It also promises a generalization to intermediary trees when data are not decisive as to their choice of tree, and a new way of constructing Bayesian priors on tree space.

This is joint work with Persi Diaconis.


Date: Thurs., Jan. 14 
Time: 4:15-5:30PM
Place: Cordura 100

Return to the seminar schedule