# Choose Ranking Functions, Customer Value Metrics, and Predictive Models A global e-commerce team asks three connected fundamentals: distinguish SQL ranking functions, define a high-value customer for a business decision, and compare linear regression, logistic regression, random forests, and gradient-boosted trees. ### Constraints & Assumptions - Ranking examples can contain ties and require deterministic output. - A high-value definition must specify horizon, margin or revenue, retention, and refunds. - Model choice depends on whether the target is continuous or binary. - Historical customer behavior can leak future information if features use the outcome window. ### Clarifying Questions to Ask - Should tied rows share a rank, and may ranks contain gaps? - What action will the high-value segment trigger? - Is interpretability, probability calibration, or raw predictive accuracy the main model requirement? ### Part 1 — Ranking functions Explain ROW_NUMBER, RANK, and DENSE_RANK with a tied example and state the extra ordering needed for deterministic ROW_NUMBER. #### What This Part Should Cover - Tie behavior and rank gaps - Partition and order definitions - A unique tie-break for deterministic selection ### Part 2 — High-value customers Define a label or score that reflects durable economic value rather than recent order count alone. #### What This Part Should Cover - Contribution margin, horizon, and refund treatment - New-versus-existing customer comparability - A validation plan tied to the downstream action ### Part 3 — Model families Compare linear and logistic regression with random forests and gradient boosting, then propose a leakage-safe evaluation. #### What This Part Should Cover - Target and loss compatibility - Linearity, interactions, calibration, and regularization - Temporal split and actionable metrics ### What a Strong Answer Covers - Definitions with examples - Economic and statistical alignment - Leakage and calibration controls ```hint Start from the downstream decision Ranking semantics, customer value, and model evaluation all change when the output is used for one winner, a tiered list, a budget-limited campaign, or an estimated probability. ``` ### Follow-up Questions - How would you evaluate value among customers too new for a long horizon? - When is DENSE_RANK preferable to RANK?