Constructing Prediction Trees from Data: The RECPAM Approach

Growing trees from the data is presented as a general way of solving the prediction problem for an unknown parameter of a distribution. A tree- structured prediction model is proposed and a strategy for building such a model from the data is presented. The strategy comprises three steps: i)RECursive partition;ii)Pruning;iii)AMalgamation; hence its acronym RECPAM. The construction is based on an information measure, the role of which is highlighted. It is shown that virtually all the available tree-growing approaches are particular cases of the general strategy.