#relative-error

Question 1: Single choice

Why is regressing $y = \log(\text{price})$ usually better than regressing the price itself with additive Gaussian noise?

Taking logs removes outliers from the data, so the fitted weights become unbiased1
Prices are positive and vary by percentages, which become additive errors on a log scale2
A log transform makes the relationship between the features and the price linear in every case3
Additive Gaussian noise overflows numerically when the targets are as large as house prices4