Can you formalize the threat model a bit more? What is the harm you’re trying to prevent with this predictive model of whether a user (new or not) will be “productive” or “really bad”? I’m mostly interested in your cost estimates for false positive/negative and your error bars for the information you have available. Also, how big is the gap between “productive” and “really bad”. MOST users are neither—they’re mildly good to mildly bad, with more noise than signal to figure out the sign.
The bayesean in me says “use all data you have”, but the libertarian side says “only use data that the target would expect to be used”, and even more “I don’t believe you’ll USE the less-direct data to reach correct conclusions”. For example, is it evidence of responsibility that someone deleted a bad comment, or evidence of risk that they wrote it in the first place?
I DO strongly object to differential treatment of new users. Long-term users have more history to judge them on, but aren’t inherently different, and certainly shouldn’t have more expectation of privacy. I do NOT strongly object to a clear warning that drafts, deleted comments, and DMs are not actually private, and will often be looked at by site admins. I DO object to looking at them without the clear notice that LW is different than a naive expectation in this regard.
I should say explicitly: I have VERY different intuitions of what’s OK to look at routinely for new users (or old) in a wide-net or general policy vs what’s OK to look at if you have some reason (a complaint or public indication of suspicious behavior) to investigate an individual. I’d be very conservative on the former, and pretty darn detailed on the latter.
I think you’re fully insane (or more formally, have an incoherent privacy, threat, and prediction model) if you look at deleted/private/draft messages, and ignore voting patterns.
Can you formalize the threat model a bit more? What is the harm you’re trying to prevent with this predictive model of whether a user (new or not) will be “productive” or “really bad”? I’m mostly interested in your cost estimates for false positive/negative and your error bars for the information you have available. Also, how big is the gap between “productive” and “really bad”. MOST users are neither—they’re mildly good to mildly bad, with more noise than signal to figure out the sign.
The bayesean in me says “use all data you have”, but the libertarian side says “only use data that the target would expect to be used”, and even more “I don’t believe you’ll USE the less-direct data to reach correct conclusions”. For example, is it evidence of responsibility that someone deleted a bad comment, or evidence of risk that they wrote it in the first place?
I DO strongly object to differential treatment of new users. Long-term users have more history to judge them on, but aren’t inherently different, and certainly shouldn’t have more expectation of privacy. I do NOT strongly object to a clear warning that drafts, deleted comments, and DMs are not actually private, and will often be looked at by site admins. I DO object to looking at them without the clear notice that LW is different than a naive expectation in this regard.
I should say explicitly: I have VERY different intuitions of what’s OK to look at routinely for new users (or old) in a wide-net or general policy vs what’s OK to look at if you have some reason (a complaint or public indication of suspicious behavior) to investigate an individual. I’d be very conservative on the former, and pretty darn detailed on the latter.
I think you’re fully insane (or more formally, have an incoherent privacy, threat, and prediction model) if you look at deleted/private/draft messages, and ignore voting patterns.