A Robust Real Time Bidding Strategy Against Inaccurate CTR Predictions by Using Cluster Expected Win Rate
Wen-Yueh Shih, Hsu-Chao Lai, Jiun‐Long Huang · IEEE Access · 2023
Previous real-time bidding (RTB) strategies offer a bidding price for each incoming bid request based on its individual predicted click through rate (CTR). However, this pricing mechanism could be a pitfall because the large and sparse feature space often leads to inaccurateindividualCTR predictions. Furthermore, our observations in a real-world online advertising environment indicate that the predicted CTR could be uncorrelated to the empirical CTR. In this paper, we introduce a new evaluation metric,cluster expected win rate (CEWR), and propose a novel frameworkCluster-aware Ranking-based Bidding Strategy (CARBS)that leverages CEWR to cope with the above issue. CEWR quantifies the worthiness of each bid request based on agroupof bid requests having similar expected performance. First, a two-step clustering method aggregates bid requests with similar predicted CTRs into clusters to gather similar information. Second,CARBSranks the clusters and sets theAffordability Thresholdin order to spend budgets smartly. CEWR summarizes the above results and hence better correlates to the click performance in our observations, causing the robustness superior to the inaccurate individual CTR predictions. Finally, a reinforcement learning-based bidding strategy is conducted to adjust thebid request expected win rate (BEWR)jointly based on CEWR and the dynamic market for deriving the final bid prices. The experimental results on three real ad campaigns manifest thatCARBSoutperforms state-of-the-art bidding strategies in terms of click acquisition. In a poorly predicted campaign (AUC: 0.73) with an extremely tight budget, the improvement is 32.5%, showing the robustness ofCARBS. The code to reproduce our results is publicly available on GitHub: https://reurl.cc/0ZoVEl.