使⽤遮蔽森林預測實時競價廣告之得標價

Abstract

實時競價廣告 (Real-Time Bidding, RTB) 在近幾年改變了網路廣告產業的運作模式。而其中，如何幫助需求方平台 (Demand-Side Platform) 在 RTB 中獲利，是許多研究者探討的主題。過去的研究中，通常將最後得標價假想由一個機率分佈所產生，但由於 RTB 本身只有得標者能獲知最後得標假的特性，DSP 所擁有的資料是一個缺損的、部分無法觀測的分佈。本篇研究將著重在如何從部分無法觀測的分佈，還原出原本的得標價分佈。基於這些因素，提出一個新的模型：遮蔽森林 (Shaded Forest)，來處理 RTB 這類部分截斷的資料。從實驗結果來看，本篇研究提出的遮蔽森林具有很好的得標價預測能力，而當無法觀測的資料比例增加時，也能有穩定的表現，並不會因此而準確率下降。

Real-Time Bidding (RTB) has changed a game changer of online advertisement. In RTB, many researchers have focus on how to maximize the profit of Demand-side platform (DSP). These researches usually consider that winning price can express as a probability distribution. However, in RTB, if a DSP lose in an auction, it will not know the winning price of that bid. Which means, what DSPs own in their data base is a partial unobserved data. In this research, we will focus on how to recover the original distribution from partial unobserved data. We propose a new model, Shaded Forest, to deal with this kind of partial unobserved data in RTB. The results of experiment show that shaded forest the accuracy of predicting winning price is better than other algorithms and have good ability to handle data with high percentage of truncation.