SHAP Values (SHapley Additive exPlanations)

SHAP values (SHapley Additive exPlanations) are a method for explaining what a machine-learning model did on one specific record. The model outputs a number — a churn probability, a lead score, a predicted order value — and SHAP breaks that single number into a set of contributions, one per input feature. Each contribution is signed: a feature that pushed the prediction higher gets a positive value, a feature that pulled it lower gets a negative one, and the contributions plus a baseline add up exactly to the prediction. That additive property is not an accident. SHAP is grounded in Shapley values, a result from cooperative game theory that fairly divides a payout among players by averaging each player's marginal contribution across every possible ordering of the others.

For marketers running propensity models, SHAP turns a black-box score into something a human can read and act on. Suppose a model flags an account as likely to churn. The SHAP values might reveal that ninety days since last login and two recent support tickets drove most of the score, while a healthy contract value slightly offset it. That is a story a customer-success manager can use tomorrow, and it beats a bare probability. SHAP works after training on any model — gradient-boosted trees, neural nets, logistic regression — and it produces both local explanations for one record and, by aggregating across records, a global picture of which features matter most across the whole population.

SHAP is often compared with LIME (Local Interpretable Model-agnostic Explanations), and the difference is worth pinning down. LIME explains one prediction by perturbing the input, watching how the model reacts, and fitting a simple linear surrogate model that approximates the black box in the small neighborhood around that record. It is fast and intuitive, but the explanation is local and can be unstable — run it twice with different perturbations and the attributions can shift. SHAP instead computes contributions with the Shapley formulation, which satisfies fairness axioms LIME does not guarantee, so the same record yields consistent attributions and the parts genuinely sum to the whole.

The trade-off is cost. Exact Shapley values require evaluating the model over every subset of features, which explodes combinatorially, so general-purpose KernelSHAP approximates them with many model calls and can be slow on large models. The practical answer is TreeSHAP, a fast exact algorithm for tree ensembles like XGBoost and LightGBM, which is why SHAP is so common on the tabular, tree-based models marketing teams actually deploy. Choose LIME when you want a quick, rough local read and speed matters most. Choose SHAP when you need consistent, additive, defensible attributions — especially for stakeholder reporting, model debugging, or any setting where an explanation might be questioned.

Use SHAP where a score has to be trusted and acted on. Build local explanations into the tools people use — put the top three SHAP contributions next to each flagged account so the human sees the reasoning, not just the number. Aggregate SHAP values across the population for a global view that tells you which features the model leans on, then sanity-check that against domain knowledge. If a leaky or nonsensical feature dominates, SHAP has just caught a modeling bug before it reached production. Match the explainer to the model too: reach for TreeSHAP on tree ensembles and reserve slower KernelSHAP for models without a fast path.

Two cautions keep SHAP honest. First, SHAP explains what the model learned, not why the world works that way — a large positive contribution shows the feature drove the score, not that it causes the outcome. Correlated features can split or trade credit in ways that mislead if you read them as causation. Second, resist over-reading tiny contributions on a single noisy record. SHAP is a lens on model behavior; pair it with the model's actual accuracy and with holdout tests before you let an attribution change a decision.

SHAP (SHapley Additive exPlanations), introduced by Lundberg and Lee in 2017, applies cooperative-game-theory Shapley values to attribute a prediction fairly across features.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where shap values (shapley additive explanations) is a core concern: