We consider dynamical systems with a linear fractional representation involving parametric uncertainties which are either constant or varying with time. Given a finite-horizon input-state or input-output trajectory of such a system, we propose a numerical scheme which iteratively improves the available knowledge about the involved constant parametric uncertainties. As its key feature, strong theoretical properties, including a structural invariance of the uncertainty's description, are preserved during the data-based learning process. In particular, it facilitates any robustness analysis and robust controller synthesis by improving the guaranteed performance. Our technique can be viewed as a data-dependent preprocessing step which supplements and enhances some recent direct data-based analysis or design approaches.