ArrayCGH data provides information on the copy number changes occurring during tumor progression. A major goal of arrayCGH data analysis is the discovery of relevant aberrations which can be used as genetic markers for predicting tumor stage, expected survival time of the patient or can help selecting suitable therapies. However, few algorithmic solutions for selection of significant aberrations are available to this date. The method we propose uses classical machine learning algorithms for feature selection to identify a subset of aberrations which has high predictive value with respect to standard clinico-pathological markers such as tumor grade, metastasis, lymph node invasion, etc. We present the results of our method when applied to a breast cancer arrayCGH data set.