Recursive partitioning (RP) is a predictive approach with minimal statistical or model assumptions, which models the relationship in terms of trees or dendrograms. It is particularly appropriate for initial exploration of large data sets, especially messy ones, and may either validate other approaches with stronger model assumptions or lead to a final analysis in its own right. It may be used for either a nominal scale (categorical) or an interval-scale (numeric) dependent variable. A major issue is the size of tree to be fitted; different approaches for this have been proposed. Like many other feature selection methods, RP is unstable, the model being potentially sensitive to minor perturbations in the data. Fitting multiple trees helps explore alternative models, and also provides better predictions than those giving by a single tree. RP is well supported in both commercial and public domain software.
|Original language||English (US)|
|Number of pages||6|
|Journal||Wiley Interdisciplinary Reviews: Computational Statistics|
|State||Published - Nov 1 2009|