The dataset has 5 labels A,B,C,D,E
A,B,C are majority and D&E are minority.The penalty to misclassify D&E are huge. How can I implement a cost sensitive learning. The input to the model will be english sentences like user comments.
3
-
$\begingroup$ Can you put up a ratio or percentages of the class representation in your data? $\endgroup$– Himanshu RaiJul 3, 2017 at 7:38
-
1$\begingroup$ Just multiply the loss for samples from D,E by their relative importance to A,B,C. $\endgroup$– EmreJul 3, 2017 at 8:34
-
$\begingroup$ A: 57% B: 17% C:18% D:5% E: 3% $\endgroup$– khangarothJul 3, 2017 at 9:21
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Option 1: play with the thresholds for classifying something – perhaps you could set a threshold for classifying something as a minority class should even if it doesn't have the highest probability/score, or a ratio of probabilities/thresholds.
Option 2: upsample your data (take resamples with replacement of the minority classes until they have as many observations as the others).
Option 3: use some cost-sensitive algorithm (e.g. some forms of classification trees) - probably won't translate into very accurate results.
