The theory of symbolic objects is used to define a notion of data that generalizes the one typical of classical data analysis (data-as-matrices). The prediction problem for symbolic objects is defined: it is seen to be a generalization of the prediction problem for classical data. An algorithm of tree-growing is developed for symbolic objects. The new algorithm is presented as a procedure for extracting knowledge from data of a more general type than classical data: it reduces to the classical algorithm of Recursive Partition when working with symbolic objects that represent classical data. Application to data consisting; of probabilistically imprecise measurements is discussed and an example is treated in detail.