An assessment asks you to “train a regressor on imbalanced data and evaluate it with precision and recall.” Explain why that request is internally ambiguous and how you would resolve it before modeling. Give a complete plan for each plausible interpretation: a rare-event classification target and a genuinely continuous regression target. ### Constraints & Assumptions - Do not relabel a continuous outcome solely to make precision and recall available without discussing information loss. - Any resampling must occur only within training data, never before the validation split. - The evaluation must match the decision the model will support. ### Clarifying Questions to Ask - Is the target categorical, binary, ordinal, or continuous? - What real-world action follows a positive prediction or a numeric estimate? - What are the costs of false positives, false negatives, and large numeric errors? - Does “imbalanced” refer to rare classes or a skewed continuous target distribution? ```hint Start with the estimand Model family and metrics follow from the target and decision; the words “regressor” and “precision” cannot both define the task without clarification. ``` ### What a Strong Answer Covers - Identification of the regression/classification mismatch. - A rare-event classification workflow with suitable splits, thresholds, and metrics. - A continuous-target workflow with regression losses and tail-aware diagnostics. - Leakage controls for weighting, resampling, tuning, and model selection. - A recommendation tied to operational costs rather than a single default score. ### Follow-up Questions - When is average precision more informative than ROC AUC? - How would probability calibration change the threshold decision? - How can a regression model be evaluated specifically on rare high-value outcomes without binarizing the target?