machine learning - Extrapolating precision/recall metrics to real-world distribution -
i've been asked report classification metrics extrapolating them real world distribution have no idea how so. here of details of problem. i've 30k records in total. out of these, run algorithm classify these 30k records 3 classes : a) correct, b) incorrect , c) unsure.
correct count : 26387 incorrect count: 1101 unsure count: 2512
out of these classes, randomly sample ~ 460 records , manually annotate create gold set. after human judgment, following counts:
correct count : 741 incorrect count: 541 unsure count: 103
i can compute precision/recall metrics, how extrapolate them real world. colleague says can consider percentage obtained classifier output metric real world scenario. mean ?
Comments
Post a Comment