PUROS: A Cpx-Ievered Framework for Ad Procurement in Autobidding Worlds

Nowadays online advertisers acquire ad impressions from digital advertising channels through operating on their au-tobidding platforms: advertisers express high level ad performance constraints for ad procurement by specifying levers including budget and cpx (e.g. cpm: max cost-per-mille, cpc: max cost-per-click, cpa: target cost-per-action). Ad platforms procure impressions on behalf of advertisers soon afterwards, and feedback ad performances (e.g. cost) to advertisers. In practice, no procurement details but aggregated ad performences are access to advertisers, and they are usually unstable because of the non-stationary autobidding environment. As a result, it's challenging for advertisers to optimize levers by directly using 'Predict, then Optimize (PO)‘ Framework (i.e. use machine-learning model with history data to estimate ad performances, then optimize levers by solving the multi-constraint optimization problem), because poor estimation performance leads to unreliable lever so-lution and results in bad performances on ad procurement. Thus, online-learning-based frameworks are recommended for lever optimization, with large amounts of history data neglected. To upgrade both the estimating & optimizing procedure in PO for better use of history data, 'Predict with Uncertainty estimation, then RObust Satisficing’(PUROS), a cpx-levered framework is presented to handle the chain reaction of poor estimation performance. In particular, an uncertainty-aware forecasting model is used for cpx-oriented ad performance estimation. Then robust satisficing is applied in the optimizing procedure, so that cpx solutions with better Pareto efficiency and robustness are produced compared to those generated by PO. Experiment results on both open-source and real-world e-commerce datasets prove the effectiveness of PUROS over traditional PO ones. Further, PUROS performs better than the previous SOTA online-learning-based framework, and shows huge potential on performance improvement with careful use of history data.