I wonder if I should exclude outlier (legit data, not wrong readings) data from my dataset using gradient boosting.

Let's say we try to predict water damage for regular houses and 99% of data is in 0-1000\$ range (even 90% is below 500\$). But there are some industrial pipes in rare cases, where the damage is 10-15k \$. It is very unlikely, that you will ever use this model to asses a high damage like that.

So from the business side, you would exclude since it is unlikely you will use the model for it. But what about the mathematical side? Will it deteriorate the model by the residuals? When the model calculates the errors, the outlier data will more likely to have a higher residual so it might become overweighted in the learning.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.