Question

A data scientist is training a regression model to predict delivery times for a logistics company. The training dataset contains a small number of extreme outliers caused by rare weather events and vehicle breakdowns. Using mean squared error (MSE) as the training loss, the model's predictions are being heavily skewed by these outliers, but the team does not want to discard the outlier records because they represent real (if uncommon) events. Which change to the training objective would best reduce the outliers' influence while still learning from them?

Accepted Answer

Switch the training loss to Huber loss, which behaves quadratically for small errors and linearly for large errors

Answer

Continue using MSE but increase the number of training epochs so the model averages out the outliers

Answer

Switch the training loss to a hinge loss to make the model robust to large residuals

Answer

Add more polynomial features so the model can fit the outlier points more precisely