Shapley values

Gold definitionUpdated Apr 2, 2026

Definition

Shapley values, originating from cooperative game theory, provide a fair way to distribute the total gain among players based on their marginal contributions. In machine learning, they are adapted to attribute the contribution of each feature or input to a model's prediction or to an outcome like an outlier.

At a glance

Executive summary

Shapley values are a fair way to determine how much each part contributes to a total outcome. In AI, they help explain why a model made a decision or what caused an anomaly, by assigning a contribution score to each input feature.

TL;DR

Shapley values fairly measure how much each input feature contributes to an AI model's output or a specific event, helping us understand its decisions.

Key points

Quantifies individual contributions by averaging marginal contributions across all possible coalitions.
Solves the problem of fairly attributing importance or causation in complex systems.
Used by researchers and ML engineers in explainable AI, anomaly detection, and causal inference.
Unlike simpler methods (e.g., feature importance from tree models), Shapley values satisfy strong fairness axioms.
Research trend focuses on developing scalable approximation methods for high-dimensional data and integrating with causal models.

Use cases

Explaining individual loan approval decisions by attributing the impact of financial features (income, credit score) to the outcome.
Identifying the root causes of performance degradation in cloud services by attributing contributions of various microservices or infrastructure components.
Determining which patient characteristics (age, symptoms, medical history) most influenced a diagnostic AI's prediction for a specific patient.
Attributing the impact of different marketing channels on a customer's purchasing decision in supply chain analytics.
Explaining why a particular transaction was flagged as fraudulent by a fraud detection system, highlighting the contributing features.

Also known as

Shapley value, Shapley feature importance, Shapley attribution