Incremental action value estimates how much better or worse a particular alternative action would perform than the observed or reference action. It supports counterfactual evaluation: instead of asking whether a prediction matched a known label, it asks how another decision policy might have performed. section 5.4
Design the metric around differences between alternatives, not merely factors that predict engagement overall:
- Identify domain factors that make competing actions meaningfully different. section 5.4.3
- Compare those factors with historical outcomes at each decision point. section 5.4.3
- Weight factors according to their ability to distinguish the outcomes of alternatives. section 5.4.3
- Validate propensity scores because inaccurate action probabilities can skew comparisons, weights, and resulting policy choices. section 5.7.2
Movie recommendations
For a horror fan, genre can create a large incremental gap between a horror and a comedy recommendation. A small rating difference, such as 4.2 versus 4.4 stars, may matter much less. Historical checks can test whether genre, popularity, language, release year, or rating differences explain clicks, watches, or completion; scoring can then amplify meaningful contrasts, such as a large genre-preference boost. section 5.4.3
