HackerTrans
TopNewTrendsCommentsPastAskShowJobs

aznicht

no profile record

comments

aznicht
·5 ปีที่แล้ว·discuss
I wonder what do you expect to hear? I don't have anything to do with youtube or its moderation approaches. But I had some experience with automatic content moderation. Imaging a system contains of tens (hundreds) of signal sources. Signal might be as complex as ML model resolution or as simple as static list of stopwords. And everything in between. With sufficiently large rate of new content to moderate (and I guess YT has a lot), you don't deal with specific moderation decisions. For you they are just points in your metrics. So you can't treat all blocks as probably false-positive, you react only to drastic change in trends. In such system the only way to solve single false-positive - human intervention. So you have some operators who process appeals. But human operators are not cheap, so first you try to increase their productivity - for example by preprocessing data for them. So they (for example) don't need to watch the video, they get transcript, probably part of it, that triggered the signal. So it might be possible that these 15 minutes were enough for human operator to make a decision based on some substring of transcript, which without context was enough for them to mark author as fraud. Or you can go even further - you can skip operator if content got enough bad signals.

My point here is that the problem is not entirely technical. You can't make ideal model without false-positive. You just try to optimise it by using discovered cases like this as input to bake better one. So the only thing to blame here - not enough human operators. Thats why you probably don't get any answers here from google engineers. The root cause either not technical, or too technical to comment it outside.