Binary tree for complex supervised learning

DETAILED DESCRIPTION As interest in complex human disease increases, there are increasing needs for methodologies that address issues such as gene-gene and gene-environment interactions in a robust fashion.
The present invention provides a binary tree-structured classification tool that addresses such needs, particularly in predicting a complex human disease (hypertension) from single nucleotide polymorphisms (SNPs) and other variables.
With a superior ability to handle combinations of predictors, the methodology disclosed herein extends beyond the traditional approach known as CART.
RTM.
, while retaining CART.
RTM.
's simple binary tree structure.
CART refers to Classification And Regression Trees where typically each split is defined by a single variable.
CART.
RTM.
is a software package from Salford Systems.
According to an aspect of the invention, the methodology includes transforming categorical predictors to indicator variables, suitably scoring outcomes, backward selecting predictors from the least to the most "important," and constructing models while respecting family structures in the data.
We successfully convert a problem of classification to one of regression without losing sight of classification.
More specifically, with our methodology, each vector of predictor values can be located to one of several disjoint subgroups, that is, to a terminal node of a binary tree.
Finding groups with high risk of disease may lend understanding to etiology and biological mechanism, as will be explained hereinafter in a later section.
In traditional classification trees, each split is on one feature alone.
What is more, the set of features is not reduced in size before any splitting is attempted.
Neither seems appropriate for polygenic disease, when no single gene is decisive and the "main effect" may be a gene by environment interaction; but most genes are irrelevant to the signal.
Algorithm Our approach has the strength of both classification trees and regression.
Although some underlying techniques are known, it is believed that our approach as a whole is utterly novel