Propensity-logged benchmarks for session-level e-commerce intervention
policies (six actions including a do_nothing control), with exact logged
propensities, counterfactual potential outcomes, multi-seed protocol, and a
baseline leaderboard.
Public logged-bandit datasets (Open Bandit Dataset) evaluate item
recommendation. Intervention policies — margin-sensitive, fatiguing,
timing-dependent actions — had no public… See the full description on the dataset page:
https://huggingface.co/datasets/tanzilidrisi/tib-benchmark.
