Metric Attribution in Internet Platforms: Concepts, Methods, and Case Studies

Metric attribution aims to identify the core factors causing fluctuations in business metrics, helping teams pinpoint why a metric rises or falls and assess the impact of each factor.

The process consists of three steps: clearly define the problem, analyze and locate the cause, and finally devise a solution.

Basic methods start with indicator judgment to confirm whether a change is genuine, followed by deterministic approaches (such as indicator decomposition using addition, subtraction, multiplication, and division) and probabilistic approaches (leveraging machine‑learning models, SHAP values, and causal inference techniques).

Deterministic judgment often uses indicator decomposition to break down a metric into contributions from sub‑metrics or conversion stages, allowing quantification of each factor's impact.

Probabilistic judgment applies modeling—typically machine‑learning combined with SHAP—to evaluate feature importance and understand how factors influence the target metric, even when causal relationships are needed.

Two case studies illustrate these methods: the first uses deterministic decomposition to identify conversion stages D and B as key drivers of a revenue decline; the second employs machine‑learning and SHAP to reveal factor A as the dominant influence on a 6.7% increase in user activity.

The article also introduces a tool that guides users through a Q&A‑style analysis, prompting them to choose decomposition or drill‑down actions and delivering corresponding insights.

Overall, the content provides a comprehensive framework for metric attribution, combining mathematical techniques and modern AI‑driven analysis to support data‑driven decision making in internet platforms.