Description
Attributes a predictive model's outputs (for example churn risk) to input features using Shapley values, showing which experiences drive predictions.
Cited in the treatise's bibliography, SHAP computes each feature's contribution to an individual prediction based on Shapley values from cooperative game theory, satisfying properties such as local accuracy and consistency. Applied to UX, teams fit predictive models of churn, conversion or satisfaction from behavioural and experience features, then use SHAP to see which experiences push predictions up or down overall and for individual users. The crucial caveat is that attributions are predictive, not causal: they explain the model, not the world, and correlated features share credit unpredictably. SHAP findings should generate hypotheses for causal methods or experiments, not substitute for them.
Attributes
- Discipline
- Computational & Machine Learning
- Loss type
- Unmigrated
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Weeks
Steps
- Define the outcome and assemble behavioural and experience features measured before it.
- Train and validate a predictive model with held-out data.
- Compute SHAP values globally (summary plots) and locally for representative users.
- Examine dependence plots for non-linear effects and interactions.
- Flag correlated features and avoid causal language.
- Turn top drivers into hypotheses for experiments or qualitative study.
Sources
- Lost Methods in UX — A Complete Treatise (Gopinath)
- Lundberg SM & Lee S-I (2017) A unified approach to interpreting model predictions. NeurIPS
- Molnar C (2022) Interpretable Machine Learning
Also known as
- SHapley Additive exPlanations
- Feature attribution for UX models
Categories
- Analytics & Behaviour
- Computational & Machine Learning