Publications

ワークショップ (国際) Leveraging Context-dependent Click Model for Off-Policy Evaluation of Ranking Policies

Haruka Kiyohara (Tokyo Institute of Technology), Nobuyuki Shimizu, Yasuo Yamamoto

The 16th ACM Recommender Systems Conference Consequences Workshop (RecSys Consequences Workshop)

2022.9.18

We leverage context-dependent click models to achieve a better bias-variance tradeoff compared to the existing estimators. Specifically, we assume that the click model variable, which is sampled conditional on the context, determines the relevant positions in the ranking. This formulation enables a more generalized click modeling compared to the existing work. Then, we propose the Generalized IPS (GIPS) estimator, which adaptively reduces the combinatorial action space depending on the click model. The proposed estimator is unbiased under any given click models and achieves the minimum variance among IPS-based unbiased estimators. The empirical results demonstrate that GIPS achieves a favorable bias-variance tradeoff compared to the existing estimators on various user behaviors.

Paper : Leveraging Context-dependent Click Model for Off-Policy Evaluation of Ranking Policies新しいタブまたはウィンドウで開く (外部サイト)