Reversing Log-transformed target after training : r2 score interpretation

Question

I have been running a log-transform on my target values because the distribution appears to be highly right skewed as you can see in the picture.

After having called

df['target'] = np.log(df['target'])

the distribution target looks like this

that is way better than before for training a model.

At this point I run the ML process and I train my model on log-scaled targets getting the following predictions (still using the log-scaled targets):

obtained by simply plotting log scaled predictions against log scaled true values, where the red line is the 'ideal' linear relationship between predictions and targets that I'm trying to achieve.

I got an R2 score of 0.40 which is not amazing but is not too bad at the moment.

The problem is, that when I try to get back to the original values by an inverse transform, i.e.

preds = np.exp(model.predict(X_test))
y_test = np.exp(y_test)

then I get the following:

and a R2 score of -0.090 obtained by running

r2_score(y_test, preds)

(hence using the inverse transformed values).

What am I doing wrong?

thank you in advance,

James

Answer 1

You should be calculating r2_score on the test data that is transformed in the same manner as the train data. In your case, the data that is transformed with log transformation, as the weights that are learned in the training part work with the transformed data.

Therefore, no need to do: preds = np.exp(model.predict(X_test)) y_test = np.exp(y_test)

Run the s2_score on preds and y_test without the np.exp transformation.

Answer 2

I got the similar situation when the performance on the log-scaled target is much better than the performance on the original scale. One of the issues is that the performance on the log-scaled target is not good enough (though is much better than the original scale). I finally found a way to improve the performance on the log scale (around 0.7 in terms of r2), leading to much better on the original scale.