Comparing strategies for robust regression tree construction

This paper deals with the problem of evaluating the robustness of regression trees. A new tree structured regression procedure is proposed, whose splitting criterion is based on M-estimation methodology. This procedure depends on a tuning parameter k. An interesting feature of this proposal is that appropriate choices for k lead to trees based on least square and on least absolute deviation criteria. A Monte Carlo experiment is employed to evaluate the performances of the proposed approach both in presence and in absence of outlying observations, compared with least square and least absolute deviation regression trees.