Publications
WORKSHOP (INTERNATIONAL) Leveraging Instrumental Variables in Online Advertising Auctions : Robust Click-Through-Rate Prediction
Ryohei Emori (Keio University), Shinya Suzumura, Takahiro Hoshino (Keio University), Nobuyuki Shimizu
ADKDD 2024: the 17th International Workshop on Data Mining and Audience Intelligence for Advertising (AdKDD 2024)
August 25, 2024
Estimates of predicted CTR in online ad auctions occupy an impor- tant position, as they are used to calculate bid amounts in automated bidding and to form rankings in Ad auctions. On the other hand, predicting CTR from historical data faces difficulties such as a bias in the direction of winning ad impressions and the cold-start prob- lem, in which the prediction accuracy deteriorates for new ads and rare users. The former is addressed by using data from sources different from ad auctions, as well as Causality-based methods such as the Inversed Propensity Score Estimator and the Doubly Ro- bust method to improve the imbalance of the data. The latter, the cold-start problem, has recently been taken up as a prediction task in a more general problem setting, known as out-of-distribution, and has attracted attention for its potential for robust prediction in accordance with the user’s invariant preference assumptions and causality-based frame. Our study reports a robust CTR prediction method for OOD by levaraging the Instrumental variables method in causal inference with an online advertising auction situation. When the IVs used are valid, it could be possible to estimate predicted CTRs similar to CTRs estimated from data with random ad impressions. We utilised bid amounts as IVs in the context of online ad auctions and incorporated explicit interactions for bid amounts and other features to account for heterogeneity in the strength of winning bid amount impressions. The characteristics of our method are that it controls for the strength of the IVs’ strength to impresion while reducing biases such as confounding bias and omitted variable bias due to the nature of DML. Our robust method for predicting CTR contributes to the literature on the robustness of causal inference in addition to research on how to predict user response in online ad auctions, and is ultimately oriented towards resolving welfare losses in online ad auctions
Paper : Leveraging Instrumental Variables in Online Advertising Auctions : Robust Click-Through-Rate Prediction (external link)