You may want to not just "minimize bias", but attempt to build a casual model (do-Calculus, causality). To give an example, recommendation might be due to "att

You may want to not just "minimize bias", but attempt to build a casual model (do-Calculus, causality). To give an example, recommendation might be due to "attended the same college" or "born in the same country". And this is probably not what you might be interested in. If you build such model and control for these spurious referrals, you might be able to get better signal. (and hopefully also promote diversity of gender, nationality, age, etc)