Modern observations of galaxies imply many telescopes, many instruments, hence yield many observables. The physics of galaxies involves many components characterized by many properties. Traditionally, the morphology was used to describe and classify galaxies because it was the only available descriptor. But one parameter is not enough to describe the complexity and the diversity of galaxies. For about thirty years, several studies have used statistical clustering approaches to devise a classification based on multiple, quantitative and objective parameters. But how can we select these parameters? How can we avoid subjectivity, redundancies, incompatibilities? In this talk I show that several statistical tools are necessary to explore the data space and select objectively the right set of parameters for the intended unsupervised clustering technique to provide a reliable result.