Databricks Certified Machine Learning Professional free question 7 / 10 · Domain 2: MLOps
A team profiles a binary classifier with an inference profile and wants roc_auc_score and log_loss in the profile metrics table. Which extra input does the Databricks documentation say these two metrics require?
- A custom metric that defines the area under the curve for a window.
- A column with the predicted probabilities, given as prediction_proba_col.
- A slicing expression defined on the label column of the profiled table.
- A baseline table with the same model identifier column as the primary table.
Show the answer and explanation
Answer: B
Why this is correct
For classification models, the profile metrics table can hold accuracy_score, log_loss, roc_auc_score, and the confusion matrix, precision, recall, and F1 statistics. The documentation states that log_loss requires prediction_proba_col, and that roc_auc_score for binary and multi-class classification also requires prediction_proba_col. The accuracy metrics appear only if the profile has the InferenceLog type and both label and prediction columns are given.
Why the other options are wrong
"A baseline table with the same model identifier column as the primary table." … A baseline table is optional and helps with drift. These two metrics need the probability column.
"A slicing expression defined on the label column of the profiled table." … Slicing expressions define subsets of the data. They do not supply the probabilities that the two metrics need.
"A custom metric that defines the area under the curve for a window." … The two metrics are built in. They need only the probability column, not a custom definition.
Probability column
Log loss and ROC AUC are computed from predicted probabilities, not from class predictions.