Metric Attribution on Internet Platforms: Concepts, Methods, and Tool Applications

Metric attribution aims to identify the core factors causing fluctuations in business metrics, helping teams locate and address the root causes of changes observed in dashboards.

The process consists of three steps: (1) clearly define the problem, (2) analyze and locate the cause, and (3) propose and implement solutions.

Basic methods are divided into three categories: deterministic judgment (e.g., metric decomposition), probabilistic judgment (e.g., machine‑learning models, SHAP values, causal inference), and speculative judgment (hypothesis‑driven analysis).

Deterministic methods include additive, subtractive, multiplicative, and divisional decompositions. Additive decomposition splits a metric (e.g., revenue) into contributions from different channels; multiplicative decomposition examines conversion funnels; divisional methods handle ratio‑type metrics.

Probabilistic methods rely on modeling: standard machine‑learning models can be enhanced with SHAP to quantify feature contributions, while causal inference or Bayesian networks are used when strong causal guarantees are required.

Two case studies illustrate the approaches. The first uses deterministic decomposition to pinpoint specific funnel stages (e.g., stages D and B) that drive a revenue decline, followed by dimension‑level drilling to refine insights. The second applies a machine‑learning + SHAP workflow to explain a 6.7% increase in user activity, identifying the most influential factor (Feature A) and its positive impact.

A dedicated attribution tool guides users through a question‑and‑answer workflow, prompting choices such as metric decomposition or dimension drilling and delivering actionable conclusions based on the selected analysis path.

Overall, the article demonstrates how systematic attribution—whether deterministic or probabilistic—enables data‑driven decision making on internet platforms.