This paper presents a methodology to transform a problem to make it suitable for classification methods, while reducing its complexity so that the classification models extracted are more accurate. The problem is represented by a dataset, where each instance consists of a variable number of descriptors and a class label. We study dataset transformations in order to describe each instance by a single descriptor with its corresponding features and a class label. To analyze the suitability of each transformation, we rely on measures that approximate the geometrical complexity of the dataset. We search for the best transformation minimizing the geometrical complexity. By using complexity measures, we are able to estimate the intrinsic complexity of the dataset without being tied to any particular classifier.