1 c c X3 0 b b X4 1 a a X5 1 b c eval NA 1 0 1 NA Estimated Propensity Score by Logistic Regression, GBDT, Random Forest 真のPSを用いるより推定されたPSを用いることで,オフライン評価の分散が小さくなる
Causal Inference by Yasui Shota • Unbiased Offline Evaluation of Contextual-bandit-based News Article Recommendation Algorithms • Efficient Counterfactual Learning from Bandit Feedback • A Contextual Bandit Algorithm for Ad Creative under Ad Fatigue • A Feedback Shift Correction in Predicting Conversion Rates under Delayed Feedback 58