3 ms·

This is your human anti-algorithm bias showing. When it is more subjective that doesn't mean that algorithm still cannot do a better job than human. :-) You hav

This is your human anti-algorithm bias showing. When it is more subjective that doesn't mean that algorithm still cannot do a better job than human. :-) You have just assumed that it can't, because you don't know how to measure it. In other words, it's an argument from ignorance (which is not surprising - AfI is often used to justify unwarranted beliefs). It's actually quite possible that algorithms learns the actual preferences of some person better than the person himself; for example, due to exponential discounting of future, person may make hiring decision based on immediate impression and discount other warning signals, which is going to be wrong in the long run. (Downvoters, can you please explain yourself? Maybe we can have an interesting discussion. Thanks.)

astazangasta 11y ago
Sure, I'll give it a shot, although I'm not amongst your downvoters. There are two ways to build a system like this: one requires no training and uses a heuristic. That is, you come up with a formula that you think will rank candidates well, plug input variables into it, and sort candidates based on the score. Such a system cannot learn, it is fixed; it can only be improved by a human tweaking it. The fact that we have no way to measure our success is irrelevant. The other way is to build a learning system, and a learning system needs to be trained. This means there must be either a set of training data (of positive examples with specified outcomes) or an objective goal function that can be computed on the output over which an optimization can be run. Without these things (i.e., when the results are too subjective and we're unable to assign a reliable score to the outcome), machine learning is impossible, and the algorithm can never be trained to perform better than a human.
- asgard1024 11y ago
  OK, so it seems to me that you pretty much agree with my argument, but you don't like my assumption that even if there is subjectivity, there can be a learning algorithm. Is that correct? I think even for 100% subjective decisions we can construct an algorithm as follows. Take the same inputs that a person (making the decisions, for example to hire) takes as input to the algorithm (well, at least try to take in reasonable number of parameters). As an example outcome, take the subjective decision that would same person (or committee, what you have) make after a certain period of time, say 3 months (when it becomes more obvious whether the original decision was good or bad). Then you can train your algorithm (but then it will, obviously, depend on specific people involved). And I think it can possibly perform better than the person for the reason I outlined.
- taurath 11y ago
  Until, of course, there is a machine that learns what the most important things to learn are.

astazangasta 11y ago

Sure, I'll give it a shot, although I'm not amongst your downvoters. There are two ways to build a system like this: one requires no training and uses a heuristic. That is, you come up with a formula that you think will rank candidates well, plug input variables into it, and sort candidates based on the score. Such a system cannot learn, it is fixed; it can only be improved by a human tweaking it. The fact that we have no way to measure our success is irrelevant. The other way is to build a learning system, and a learning system needs to be trained. This means there must be either a set of training data (of positive examples with specified outcomes) or an objective goal function that can be computed on the output over which an optimization can be run. Without these things (i.e., when the results are too subjective and we're unable to assign a reliable score to the outcome), machine learning is impossible, and the algorithm can never be trained to perform better than a human.

asgard1024 11y ago
OK, so it seems to me that you pretty much agree with my argument, but you don't like my assumption that even if there is subjectivity, there can be a learning algorithm. Is that correct? I think even for 100% subjective decisions we can construct an algorithm as follows. Take the same inputs that a person (making the decisions, for example to hire) takes as input to the algorithm (well, at least try to take in reasonable number of parameters). As an example outcome, take the subjective decision that would same person (or committee, what you have) make after a certain period of time, say 3 months (when it becomes more obvious whether the original decision was good or bad). Then you can train your algorithm (but then it will, obviously, depend on specific people involved). And I think it can possibly perform better than the person for the reason I outlined.
taurath 11y ago
Until, of course, there is a machine that learns what the most important things to learn are.