Combination effects seem challenging as you point out. I think it’s often taken for granted that weighting things should be done linearly, but there really isn’t any reason to believe this would approximate the moral truth or what we’d want to care about upon reflection in this domain, although it’s useful for its simplicity, interpretability and transparency.
Another specific challenge is whether we should apply a given (usually monotonic) transformation to a feature that comes in degrees first. For example, if the degree of X matters, say neuron count or neuron count in a particular part of the brain, should we use X,X2,√X,logX,2X or something else? There are infinitely many degrees of freedom here.
Combination effects seem challenging as you point out. I think it’s often taken for granted that weighting things should be done linearly, but there really isn’t any reason to believe this would approximate the moral truth or what we’d want to care about upon reflection in this domain, although it’s useful for its simplicity, interpretability and transparency.
Another specific challenge is whether we should apply a given (usually monotonic) transformation to a feature that comes in degrees first. For example, if the degree of X matters, say neuron count or neuron count in a particular part of the brain, should we use X,X2,√X,logX,2X or something else? There are infinitely many degrees of freedom here.