Which Metric to consider for Regression R2 or RMSE

evaluation_metric

#1

Hi All,

We are working on Regression problem usinng Sklearn, So what should be best Metric for this R2 or RMSE or any other.

We are getting R2 in range of 0.35 , is it fine or not.

Regards,
Jai


#2

Adjusted R2 is the best metric.
Value around 0.4-0.5 is good.


#3

Hi,

RMSE is root mean squared error. It is based the assumption that data error follow normal distribution.

This is a measure of the average deviation of model predictions from the actual values in the dataset.

R2 is coefficient of determination, scaled between 0 and 1.

R^2 = 1-(SSE/SST)

SSE : sum of squared error, SST : total sum of squares

R-squared is simply the fraction of response variance that is captured by the model.

If R-squared = 1, means the model fits the data perfectly.

while both indicate the goodness of the fit, R-squared can be more easily interpreted.

It directly measures the goodness of fit in capturing the variance in training data.

For example : if R2=0.7, it says that with this model, we can explain 70% of what is going on in the real data,

rest 30% can’t be explained.

if your R2 is in the range 0.35, then model explains only 35% of the variance. This is not a good fit.

something in the range 0.7-0.8, is a good model.

I hope this helps.


#4

Nicely explained @himansu979,
One more comment regarding R2 and Adjusted-R2 : -

R2 increases as we add more number of explanatory variables into the model. Thus, to correct for this situation - Adjusted-R2 is measured in following way:

Adjusted-R2 = 1 - [(N-1)/(N-d)](1 - R2)

Where, N = number of data points. (d + 1) = number of parameters in the model.