Performance is not enough: a story of the Rashomon's quartet

Predictive modelling is often reduced to finding a single best model that optimises a selected model quality criterion. But what if the second best model describes the data equally well but in a completely different way? What about the third best? Following the Anscombe's quartet point, in this paper, we present a synthetic dataset for which four models from different classes have practically identical predictive performance. But, visualisation of these models reveals that they describe this dataset in very different ways. We believe that this simple illustration will encourage data scientists to visualise predictive models in order to better understand them. Explanatory analysis of the set of equally good models can provide valuable information and we need to develop more techniques for this task.