English   -  Français  

Résumé de communication

Résumé 293 :

Tree methods in statistical pattern recognition
Devroye, Luc
McGill University

Recursive partitions of the space can be used to classify data in a very natural manner. Statistically equivalent blocks and histograms are early examples of this. It was not until the era of CART that statisticians studied recursive partitions in a concerted and organized way. Breiman understood the power of tree-based methods and proposed a family of classifiers under the umbrella of random forests. A few recent theses were devoted just to tree methods. The recurring question is: "How does one approach the design?" In fact, what are the right questions to ask? Is it better to partition the space into very small pieces and then recombine regions (as in CART), or is a top-down design preferred? Or can both coexist? And how does avoid the design parameter conundrum that plagues most nonparametric problems?