Predicting obesity using machine learning techniques: Classification trees and random forests.

Abstract

Data Analysis: The data was separated into two random sets of each sex (training and test) for crossvalidation. Seven trees were calculated in the training group, using different numbers and combinations of predictors, using the R package tree (Ripley, 2013). The model (tree) with lower deviance and misclassification, as well as higher Pseudo-R2 was selected as the best predictive model for each sex. This model was applied in the test set to verify its accuracy. In order to avoid overfitting, a Random Forest analysis was conduced in the training set, using the R package randomForest (Liaw & Wiener, 2013). A weighted network was plotted using the proximity matrix as input, in order to visualize the quality of the random forest’s prediction. A cross-validation procedure was conduced in the test set to test the random forest classifier. A new weighted network was created, for each sex, to compare the predictive quality in the training and in the test set.