To obtain a good modeling of a phenomenon, data and methodology are as important as the modeling technique used if this technique is an inductive machine learning one such as artificial neural networks or decision trees. We are interested to build gold predictive maps using a Geographical Information System. As for any classification task, we have to discriminate patterns of a desired class from other ones. The choice of counter examples is a crucial stage for a deposit classification. The first idea to have counter examples is to define barrens. This approach introduces skews because the barrens are usually far away from deposits with very different characteristics or close enough but without any certainty that they really are barren deposits. The goal of this study is to propose a new more general subjective approach to build dataset for classification based on the main properties of geology : type of rocks, geologic time scale, geographic zone, substances,? The main idea is to consider a deposit like a pattern characterized by a certain number of properties. Then, any pattern with at least one different property can be used as a counter example. Then, we apply on a GIS Andes these various approaches to create efficient datasets aim to obtain predictive maps on gold deposits using artificial neural networks. Thus, we first try to discriminate neogene gold deposits and old gold deposits. Then, we try to discriminate neogene deposits containing gold vs neogene deposits without gold. We obtain a minimum rate of good classification of 85% using 31 attributes describing geography, geology (lithology, faulting, recent volcanism), geometry of the subduction zone, geothermy, geophysics (seismic, gravimetry), and ore deposit. Finally, we compare the different maps obtained and discuss how to combine them using e.g. a bagging method.