CODEWITHSUNDEEP
machine-learningclassificationdata-sciencethreshold
Neo
Neo

Reputation: 4478

How to decide threshold in classification model?

Suppose I build a classification model and then to improve, lets say,precision I just increase my threshold probability of higher class. Does this make sense? I am not changing the model but just changing the threshold probability to get better answer. Is it ok? Thanks

Upvotes: 3

Views: 3738

Answers (2)

Martin Thoma
Martin Thoma

Reputation: 136875

Parameter tuning, such as setting a threshold according to your data, is perfectly fine.

However, keep in mind that you should make a train-test split of your data. The training data is used to calculate your parameters, the test data should only be used once at the very end, when you want to calculate how well your algorithm performs. If you need two datasets to calculate your parameters (e.g. some parameters and then another dataset for the threshold), then split your training dataset again (now you have training-, validation-, and test-data).

Upvotes: 2

SmallChess
SmallChess

Reputation: 8156

That's perfectly acceptable, in fact that's one of the reasons why we have ROC curve and precision recall curve.

Changing threshold is known as parameter tuning and is a common practice.

Upvotes: 3

Related Questions