Multiple Trees or Forest of Random Trees Analysis - One model versus many models

›› See also Correlation/Interaction Effects
›› See also Observation Distance Matrix

While one model of the data is informative, many alternative models or trees are possible. The "greedy" approach to tree building uses the most significant predictor variable at each split. However, there may be many other highly significant splitters that can give alternate views of the data.

Figure. Multiple tree model variable list viewer.

Because the number of possible trees that can be generated is exponential in the number of significant predictors, we turn to random sampling. In multiple tree analysis, each tree is created by randomly choosing predictors among the most significant predictors.

HelixTree software's multiple tree (forest of random trees) analysis is powerful enough to generate hundreds or thousands of random trees on a data set, and then evaluate the forest of trees as a whole.

The random tree predictors view, depicted in the screen at right, allows the sorting of trees by the four columns; the three RMS error columns or the number of leaves in the tree column.

One could apply the Occam´s razor principle to select the smallest tree that has a given RMS error - that is, among models with equal predictive value, choose the simplest one.

Once a multiple tree file has been generated, there are several analysis tools at your disposal to analyze the data.