Deconfounding Candidate Support in Offline Web Behavior Prediction
November 11, 2026
Current web-behavior benchmarks suffer from a candidate-support confound where models identify target actions simply by their presence in training support. This paper introduces the support-identifiability rate (SIR) and a dual-path proposer that achieves 0.7214 Exact@1 on the OPeRA dataset by balancing support.
HOW THIS AFFECTS YOU
●
builderThis shows you how to build better rerankers that work on out-of-vocabulary user actions.
●
researcherYou can use the support-identifiability rate to ensure your behavior prediction models are actually learning instead of exploiting data distributions.