Study Databricks ML-ASSOC Classification and Regression Metrics: key concepts, common traps, and exam decision cues.
Many ML-ASSOC misses are not about the model. They are about choosing the wrong metric for the business objective or interpreting the result on the wrong scale.
| Task | Metrics that often matter |
|---|---|
| classification | F1, log loss, ROC/AUC, and related class-sensitive metrics |
| regression | RMSE, MAE, R-squared, and other error or fit measures |
| If the business mostly fears… | Stronger metric instinct |
|---|---|
| missed positives | recall-sensitive thinking |
| false alarms | precision-sensitive thinking |
| overall class ranking quality | ROC/AUC can matter |
| large regression misses | metrics that punish larger errors more strongly |
| average absolute error size | MAE often fits better |
The exam rewards choosing a metric because of the business loss pattern, not because it is the metric you see most often.
| If the stem says… | Better first instinct |
|---|---|
| imbalanced classification risk | avoid defaulting to accuracy without thinking |
| large errors are especially painful | consider metrics that penalize large errors appropriately |
| target was log-transformed | remember predictions or metrics may need exponentiation for interpretation |
If the answer sounds mathematically normal but ignores the business objective, it is usually weaker than Databricks wants.
| Trap | Better rule |
|---|---|
| choosing a familiar metric instead of the scenario metric | match the business objective first |
| forgetting the transformed scale of the target | log-transformed targets can distort interpretation if left on the wrong scale |
| celebrating one strong number without context | metric fit still depends on the scenario |
| Scenario clue | Stronger answer shape |
|---|---|
| “rare positive class matters” | do not default to accuracy |
| “false negatives are costly” | favor recall-aware reasoning |
| “large misses are disproportionately bad” | use a regression metric that reflects that penalty shape |
| “target was log transformed during training” | check whether exponentiation is needed before interpretation or evaluation |
Metric-choice questions usually start with the business mistake that matters. Accuracy can be weak for imbalance. Log transforms affect later interpretation and sometimes require exponentiation before reading predictions correctly. The exam usually rewards matching the metric to the scenario objective rather than choosing the most familiar score.