Databricks Machine Learning Associate Question 259
Single answerYou are building a regression model to predict house prices. During training, you transform the target variable (house price) by applying a natural logarithm to stabilize variance. After training and evaluating your model, you compute the Root Mean Squared Error (RMSE) directly from the predictions and true log-transformed values. What could be the issue with this evaluation process?
- A
The RMSE is calculated correctly and no additional steps are needed.
- B
The RMSE needs to be exponentiated to reflect the scale of the original target variable.
- C
The RMSE needs to be squared to reflect the scale of the original target variable.
- D
The RMSE is not affected by the log transformation of the target variable.
Show answer and explanation
Correct answer: B
Explanation
When the target variable is transformed using a natural logarithm, the model's predictions and evaluations are on the log scale. While RMSE is computed correctly on the log-transformed scale, it does not directly correspond to the original scale of the target variable. To interpret the RMSE in the context of the original target variable, exponentiation is required. Without this step, the RMSE will not provide meaningful insights into the errors on the original scale.
- A. Incorrect.
This is incorrect because when the target variable is log-transformed, the RMSE is computed on the log scale and does not directly correspond to the original scale. Additional steps are needed to interpret it correctly.
- B. Correct.
This is correct because when the target variable is log-transformed, the RMSE must be exponentiated to return to the original scale of the target variable, ensuring meaningful interpretation in the context of the problem.
- C. Incorrect.
This is incorrect. Squaring the RMSE is not the correct way to transform it back to the original scale. RMSE is already a squared-rooted metric by definition, and exponentiation is required, not squaring.
- D. Incorrect.
This is incorrect because the log transformation directly affects how errors are computed. The RMSE computed on the log-transformed scale does not automatically reflect the errors on the original scale of the target variable.