3 ms·
What I meant is that I expect a huge initial win, as this catches 80-90% of spammy edits, until the spammers figure out how to get around it. Hopefully the cons
by zellyn 11y ago
What I meant is that I expect a huge initial win, as this catches 80-90% of spammy edits, until the spammers figure out how to get around it. Hopefully the constant stream of human-created training data (in the form of reverted or non-reverted edits) helps with this. But undoubtedly there will need to be some continual feature engineering to stay ahead.
- halfak_wmf 11y agoYeah. We're actively working with patrollers to chase down mistakes and retrain the models. This is a system that will require continual improvement. Luckily, there's a lot of researchers working on this problem so we're drawing from a lot of brilliance beyond our little team at the Wikimedia Foundation.