Analytics Glossary
SHAP (SHapley Additive exPlanations) is an explainable-AI method that attributes a machine-learning model's prediction to its input features, showing exactly which factors pushed the result up or down and by how much. It is grounded in Shapley values from cooperative game theory and was formalised by Lundberg and Lee in their 2017 NeurIPS paper.
SHAP treats every prediction as a game: the input features are the players, the prediction is the payout, and each feature gets credit for its fair share of the outcome.
SHAP borrows the Shapley value from cooperative game theory — the only way to divide a payout among players that satisfies fairness axioms such as consistency and local accuracy. A feature's SHAP value is its fair contribution to the prediction.
SHAP values add up: the model's baseline (average) prediction plus the SHAP value of every feature equals the actual prediction. That makes each explanation auditable — nothing is hidden in a black-box residual.
A single prediction can be explained on its own (why was this order flagged?), and thousands of SHAP explanations can be aggregated to reveal which drivers matter most across the whole business.
An illustrative SHAP waterfall for a margin move.
The model’s expected value (base) plus each driver’s signed contribution sums exactly to the final number — bars pushing up add to the metric, bars pulling down subtract from it. That additive property is what makes SHAP attribution auditable: every point of the change is accounted for by a named driver, nothing is left in a black box.
A prediction you cannot explain is a recommendation you cannot defend. SHAP turns opaque model outputs into ranked, quantified drivers that decision-makers can question, verify and act on.
In the Explain step of Adaptrix's analytics chain (Connect → See → Explain → Predict → Act), SHAP-style feature attributions are translated into plain-language driver summaries — a KPI change arrives with the why attached, not just a chart. And because Adaptrix can run fully self-hosted, those explanations are computed without your data leaving your infrastructure.
Keep building your analytics vocabulary.
Both explain individual predictions, but SHAP is grounded in Shapley values and guarantees properties such as local accuracy and consistency; LIME fits a local surrogate model without those guarantees. The 2017 SHAP paper in fact shows LIME to be one of six methods unified under the SHAP framework.
No. SHAP explains what the model used, not what causes the outcome in the real world. If the model learned a spurious correlation, SHAP will faithfully report it. For decisions that depend on cause and effect, SHAP should be paired with causal analysis.
Not necessarily. Open-source libraries make SHAP accessible to practitioners, and analytics platforms increasingly embed SHAP-style attributions directly into dashboards, translating them into plain-language driver summaries for business users.
Exact Shapley values are exponentially expensive to compute, but practical approximations — such as TreeSHAP for gradient-boosted trees and sampling-based estimators for other models — make SHAP fast enough for production analytics on realistic workloads.
Bring a real question from your business and watch a live answer — reasoning, sources and math included — or model what it saves you first.