While considerable amount of experimental catalytic data are available in literature and databases, the inconsistency between different sources often impede the learning of accurate models. In this talk, I will introduce our recent method sign-constrained multi-task learning as implemented in the SISSO framework, termed SCMT-SISSO, for distilling accurate descriptors from experimental data in the example of predicting the catalytic activity of perovskite oxides for oxygen evolution reaction (OER). While many previous descriptors for the OER activity were proposed based on respective small datasets, we obtained the new 2D descriptor (dB, nB) with greatly improved universality and predictive accuracy based on 13 experimental datasets from different publications. This descriptor allowed us to identify hundreds of unreported highly active perovskites from a large chemical space. Experiments were performed on several of the candidates, which confirmed two new perovskites that are highly active (>BSCF5582) for OER.