Abstract:
|
Tree-Based Models, including Classification and regression trees (CART), provide a useful alternative to more standard regression techniques using linear predictors. These models are especially useful for forming diagnostic or prognostic rules. However, the predictive models obtained from trees typically involve complex, high order interactions between the modeled covariates, and it is therefore difficult to visualize the results of a tree model in a succinct manner. In particular, reports of the fitted tree may obscure similarity of response distribution among regions that are actually adjacent in Cartesian space. We present CARTscans, a graphical tool providing a view of the structure of a tree model. Predicted values are displayed across a four-dimensional subspace of the covariates, with smoothing of effects due to other covariates. Using these graphs, a user is able to take advantage of the flexibility of tree-based models to find complex interactions and pick out interesting regions while still being able to visualize main effects.
|