Should organisations use Shapley values or simpler feature importance methods for explainability?

When Shapley values are worth the extra cost

Shapley values make sense when explainability has to survive challenge, not just inform a model owner. They are useful when a decision affects a customer, patient, employee, credit decision, or safety outcome and you need an attribution method that is consistent, defensible, and less dependent on arbitrary feature order or model-specific heuristics.

That makes them a stronger fit for review processes where the explanation itself becomes part of the control, such as model governance, dispute handling, or oversight of decisions with material business or regulatory impact. They are also more appropriate when multiple stakeholders must trust the same explanation and when a “directional” explanation would be too weak to support a formal review.

Shapley-style attribution is strongest when the organisation is prepared to pay for that rigor in compute time, latency, and implementation complexity. If the use case is batch review, offline validation, or a small number of high-consequence cases, the cost is often justified by the quality of the explanation.

When simpler feature importance is the better choice

Simpler feature importance methods are usually the better operational choice when the goal is to understand which variables generally move the model, not to produce a full explanation for scrutiny. They are faster, easier to communicate, and often easier to operationalise in dashboards, monitoring, and iterative model development.

These methods are typically sufficient when the business question is directional, for example whether income, tenure, usage, or recency appears to matter more in aggregate. In that setting, the explanation supports model debugging, feature review, or rough prioritisation, and does not need to carry the burden of a formal justification for an individual decision.

They are also attractive when the model is changing frequently and the team needs feedback quickly. If the explanation will be used to guide iteration rather than to defend a decision, the extra precision of Shapley values may not change the action taken.

How to choose the right explanation method

The choice is less about which method is “better” in the abstract and more about what decision the explanation must support. If the output will be reviewed by humans, challenged by stakeholders, or documented as evidence for a sensitive decision, favour the method that is most defensible even if it is slower. If the output is mainly for model tuning, monitoring, or internal insight, favour the method that is cheapest and clearest for the team using it.

A practical rule is to match the method to the decision risk and the audience. Where the answer must be auditable, repeatable, and robust to scrutiny, use the more rigorous method. Where the answer is simply meant to orient the business toward the main drivers, a simpler method usually gives better value.

For organisations that want a control-based framing, explanation quality is part of decision integrity. That is why model governance and review processes should define, in advance, which classes of decisions require stronger attribution and which can rely on lightweight directional methods.

Practitioner Guidance

What to prioritise: Separate high-stakes decisions from operational analytics before picking an explanation method. If the explanation may be used in a dispute, approval, adverse action, or other formal review, treat attribution quality as a control requirement rather than a convenience.

Decision rule: If a weaker explanation would not change the decision, a simpler method is usually enough. If the explanation must withstand challenge or be reused as evidence, choose the method that gives the most stable and defensible attribution, even if it costs more.

What to measure: Track whether the chosen method actually changes human decisions, reduces review time, or improves confidence in model outputs. If not, the organisation may be paying for explanation detail that nobody uses.

Practitioner takeaway: Pick the simplest method that still matches the accountability burden of the decision. Explainability is only worth the extra compute when the explanation itself must be trusted as part of the decision process.

Related resources from NHI Mgmt Group

• When should organisations use OAuth instead of simpler authentication methods for agents?
• Who should use digital certificates instead of simpler MFA methods?
• When should organisations use behavioral biometrics instead of other passwordless methods?
• How should organisations use proxy methods for protected characteristics in fairness analysis?