Databricks Machine Learning Associate Question 261
Single answerYou are evaluating the performance of a regression model that was trained to predict house prices. The model was trained using the natural logarithm of the house prices as the label variable. After obtaining predictions from the model, you calculate the Root Mean Squared Error (RMSE) and notice it is unusually low. Which of the following is the most likely reason for this observation?
- A
The RMSE was computed on the logarithmic scale and was not exponentiated back to the original scale.
- B
The RMSE calculation contains a bug, and the metric was incorrectly computed.
- C
The model is overfitting to the training data, leading to an artificially low RMSE.
- D
The dataset used for evaluation contains outliers, which skewed the RMSE calculation.
Show answer and explanation
Correct answer: A
Explanation
When the log of a label variable is used for training a regression model, the RMSE is computed in the logarithmic scale. To interpret the RMSE in the original scale of the data, it must be exponentiated. Failure to do so results in a misleadingly low RMSE, as the value remains in the logarithmic scale and does not reflect the true scale of the original data.
- A. Correct.
This is the correct answer. When the log of the label variable is used, the RMSE is initially computed on the logarithmic scale. To interpret it in the original scale of the data, the RMSE must be exponentiated back. Failure to do so can result in an unusually low RMSE.
- B. Incorrect.
While there could be bugs in the RMSE computation, this is not the most likely explanation in this scenario where log-transformed labels were used.
- C. Incorrect.
Overfitting can lead to low RMSE on training data, but this does not explain the specific issue of unusually low RMSE caused by log-transformed outputs.
- D. Incorrect.
Outliers can affect RMSE, but they would typically increase the RMSE rather than lowering it. This is not related to the log transformation issue described in the scenario.