Data‑Driven Metric System Construction and Application: Theory, Methods, and Real‑World Cases

The presentation introduces the concept of a metric system, emphasizing the need to start from the business goal (the "end") and decompose it into hierarchical indicators and dimensions. Two classic examples—an e‑commerce app and a content‑type product—illustrate how basic metrics (e.g., GMV, DAU) are broken down into sub‑metrics.

Beyond pure business‑driven design, the talk stresses the importance of data‑driven metric construction. It outlines the OSM framework (Objective, Strategy, Metric) for business‑driven metrics and then introduces data‑driven methods, including the GSM framework (Goal, Signal, Metric) and the HEART model for selecting candidate signals.

Three practical development branches are described: statistical analysis (correlation, visualization), machine‑learning approaches (Shapley values, decision trees, clustering), and causal analysis (Uplift, DML). Each branch is illustrated with real cases from a video‑app, such as filtering out automatic‑play impressions and using uplift modeling to link playback completion to user retention.

The article then details the full lifecycle of metric system application: alerting (statistical rules or predictive models like XGBoost), diagnosis (attribution analysis using multi‑dimensional breakdowns and Gini coefficients), and strategy formulation (targeting user groups, scenes, and traffic). A concrete "in‑play intervention" workflow demonstrates how early warning, diagnosis, and targeted actions can improve content performance.

Overall, the content provides a comprehensive guide for product and data teams to design, validate, and operationalize metrics that are both directionally aligned with business goals and sensitive enough to detect meaningful changes.