Hacker News Reader

Comment by cs702 | Hacker News Reader

I read the intro and skimmed the paper. The OP doesn't propose to "make loans without predictive optimization." It proposes that the people who use software to make loans specify how they are dealing with these issues, instead of sweeping them under the rug.

View on HN

Exactly this. If you're going to use an algorithm to discriminate between people in order to advantage some over others (by granting a loan to them, or admitting them to a university, for example) you'd better be able to fully and independently verifiably explain how your algorithm works and doesn't discriminate among people due to their protected attributes or proxies thereof. Current machine learning technologies render this assurance all but impossible and so they must be regarded with the most extreme suspicion until this can be provably remedied. Given how difficult it is validating traditionally programmed inference systems we're a long way from it.

If a person and not an algorithm decides,there is no full and verifiable explanation. If we want that we need to move to fully algorithmic decisions to start with.

But a person can (and will, sometimes) be held accountable.

Nothing stops us from holding algorithm users accountable.

It certainly would help then, if the algorithm was scrutable.

It could be judged on outcome alone.

If you already know what the outcome is, why bother with the model in the first place?

Sorry, not the outcome of the model, but the outcome of the use of the model, i.e., the ex-post effect.

Who is holding loan officers accountable?

I agree. Also it’s worth noting that there are multiple, reasonable definitions of “fair” and separately following these will not result in equivalent outcomes. So no matter what is done, there will be room for criticism.

> due to their protected attributes or proxies thereof

By definition, if it's algorithmic, it's not discriminating due to protected (or any other) attributes - it's discriminating very strictly on ability to repay with absolutely no bias, because it's literally an unthinking, unfeeling machine. Forcing ML models to find specific results rather than actually letting the models do what they were designed and asked to do renders them useless, but that may be the goal.

We totally do discriminate based on protected attributes for university admissions though

Reply on news.ycombinator.com