How do you frame high-value customer identification as classification versus regression?

WHAT THIS TESTS: This question tests whether you can translate a fuzzy business request into a rigorous machine learning formulation. The interviewer cares about your ability to interrogate the problem definition, choose an appropriate target variable, and align the model output with a downstream business action. It also reveals whether you understand that classification and regression are not just algorithmic choices but fundamentally different ways of defining error and success.

A GOOD ANSWER COVERS: A strong answer starts by asking clarifying questions: how does marketing define value, what is the budget or reach of the campaign, and what action will be taken. Next, it frames regression as predicting a continuous quantity such as expected revenue, lifetime value, or transaction count, which naturally produces a ranked list and allows fine-grained segmentation. It then frames classification as predicting a discrete label such as a binary high-value flag or an ordinal tier like low, medium, and high, which simplifies decision-making and can directly gate a yes-or-no treatment. A great candidate also compares the trade-offs: regression preserves ordering and granularity but requires defining a continuous metric and choosing a cutoff; classification collapses nuance into buckets but aligns cleanly with fixed campaign segments and can yield cleaner precision-recall metrics. Finally, the answer should connect the choice to the campaign action, for example noting that a regression output is better when you want to prioritize the top decile by expected spend, while classification is better when you have a fixed offer for a specific segment and need a hard gate.

COMMON WRONG ANSWERS: A red flag is jumping straight to model names like XGBoost or logistic regression before defining the target variable. Another mistake is claiming that classification is always better for business because it gives a yes-or-no answer, without acknowledging the information loss from binarizing a continuous signal. Conversely, some candidates advocate for regression without explaining how the business will actually use a continuous score. Failing to mention that the campaign action determines whether you need a calibrated probability, a ranked list, or a thresholded segment also signals shallow product thinking.

LIKELY FOLLOW-UPS: The interviewer may ask how you would handle class imbalance if you choose classification, or how you would define the continuous target for regression if customer value is realized over multiple years. They might also probe how you would evaluate the two approaches offline, for example by comparing the lift curve from a regression ranker against the precision of a classifier at a fixed budget. Another common follow-up is asking how you would turn a regression model into a classification-like decision or vice versa, which tests whether you understand that the two framings live on a spectrum.

ONE CONCRETE EXAMPLE: Suppose the marketing team has a 10,000-dollar budget to send gift boxes to the most promising 1,000 customers. If you frame this as regression, you predict expected three-month spend, rank customers by that score, and mail the top 1,000. This maximizes expected return but requires you to trust the point estimates. If you frame it as classification, you define high-value as historical spend above 500 dollars, train a binary classifier, and mail everyone predicted positive. This is easier to explain to stakeholders and lets you tune precision and recall, but it ignores the difference between a 550-dollar customer and a 5,000-dollar customer. A senior answer might propose a hybrid: use regression to predict value, then apply classification-style thresholds to create tiers so that marketing can run A/B tests across segments.

Source: ibm.com