Predicting the propensity for thermally activated β events in metallic glasses via interpretable machine learning