Asymmetric feature interaction for interpreting model predictions