Is negative_mean_squared greater the more accuracy or less computes to more accuracy?

Question

-0.567 -4.235
Which of the above negative_mean_squared error value computes to more accuracy?

Answer 1

Higher the value better it is. So in your case -0.567 is better. A per the documentation sklearn scoring functions maintain the following convention higher return values are better than lower return values. But when you look at mean_squared_error or for that matter even mean_absolute_error lower the value better it is. So they just flip the sign of the value to make sure it follows the convention. For your example, -0.567 results in a mean_squared_error of just 0.567 whereas, -4.235 has a mean_squared_error of 4.235 which is much higher than the former.