Reproduce “Beyond Marginal Feature Coverage: Identifiability and Certification in Off-Policy Evaluation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Prahas Duggireddy
The authors study when logged decisions identify a target value in off-policy evaluation. They show feature-by-feature coverage is insufficient: identification holds exactly when the target feature row lies in the span of logged rows, under a finite-action, exact bounded linear reward model with an interior neighborhood. In an audit, 12–24 of 720 groups passed marginal coverage but failed the span test.
Key insights
- Marginal feature coverage is insufficient for identifiability in feature-based OPE; true identifiability requires that target features lie within the linear span of jointly logged feature vectors.
- When a target quantity is not uniformly identified, its true value is not a single point but can be characterized by a "compatible-value interval" given reward bounds and logged data.
- Existing OPE implementations, relying on weaker support rules, may unknowingly report definitive point estimates for non-identifiable quantities, fostering a false sense of precision.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.