- 學位論文
第一高價即時競價下控制預算競標策略
A First-Price Auction Bidding Strategy For Budget Control
摘要
本作的目的是為了解決廣告需求方平台對於實時競價出價的精準度隨著第一高價拍賣機制比例逐漸擴大而越發重要，而在實際的場景中，預測單筆交易的勝利出價非常困難。我們發現第一高價拍賣的行銷活動中勝率分布圖與許多先前的第二高價拍賣研究相反，在使用相同的預測目標點擊率模型與相同的出價模型及預算限制下，不同價格區間彼此之間的勝率相近，而非隨著出價提高而提升，其原因在於不同DSP間能獲得的使用者資訊大致相同，導致DSP間的預測模型精準度相差無幾，代表DSP間對於商品有類似個估值，所以在相同的預算限制下，勝率相近。我們根據上述觀察得出勝率函數只和預算限制有關，換言之出價策略在寬鬆的預算限制下會溢價導致利潤率變低，本作提出在透過機器學習的方式下平衡收益及利潤率的第一高價拍賣出價的模型
並列摘要
This paper aims to solve the problem that the accuracy of real-time bidding price on the advertising demand-side platform is becoming more important as the proportion of the first high-price auction mechanism gradually expands. In actual scenarios, it is not straightforward to predict the winning bid of a single transaction. We found that the distribution of the winning rate in the marketing campaign of the first-highest price auction is contrary to many previous studies on the second-highest price auction. The winning rates of different price range under the same click-through rate model, the same bidding model, and the same budget constraint are almost identical. The observed phenomenon is because the user information obtained between different DSPs is roughly the same, resulting in the same accuracy of prediction models between DSPs. Therefore, under the same budget constraint, the winning rate is similar. Based on the above observations, we conclude that the winning rate function is only related to the budget constraint. In other words, the margin of the DSP is decreasing when setting loose budget constraints. This work proposes a machine learning method in the first high-price auction bidding model that balances revenue between the margin.