Loading…
6
My buddy Dave in Denver said just throw more data at my model, so I did and it got worse
I was training a small model to sort support tickets and it kept missing the refund ones, so Dave (who does ML at a startup) said double the training data and it would fix itself. I scraped 40,000 more labeled tickets over two weekends and the accuracy on refunds actually dropped from 78% to 71%. Turned out like 30% of the new labels were junk because I outsourced them cheap, so I basically taught the thing to be confused. Anyone else get burned adding data without checking if it was any good first?
0 comments
Log in to join the discussion
Log In0 Comments
No comments yet
Be the first to share your thoughts on this discussion.