在即時競標系統利用伽瑪分佈預測廣告成交價

Abstract

在廣告即時競標系統(Real-Time Bidding)中，預測廣告成交價以用來評估競標花費成本及決定競標價錢對DSP (Demand-Side Platform)的競標策略而言是一個相當重要的問題。目前的方法是利用設限線性回歸(censored linear regression)同時考慮贏的和輸的競標資料在預測廣告成交價上。傳統的線性回歸模型是以常態分佈為基礎，但是廣告成交價只會是正值，而且很難將常態分佈和廣告成交價之間用物理意義聯想一起。因此，我們認為常態分佈對廣告成交價的預測並不適合。在這篇論文中，我們提出了以伽瑪分佈為基礎的線性回歸模型(Gamma-based linear regression)，此模型也同時考慮贏的和輸的競標資訊。為了從競標資訊中找出我們模型的參數，我們採用兩階段的方法，讓我們可以簡易地得到我們需要的參數，除此之外，為了加快收斂的速度，我們也設計多種啟發式(heuristic)方法用於初始參數的設定。從實驗結果顯示，我們的方法在兩個實際的資料中都比目前的方法更適合用來評估廣告成交價。

In Real-time Bidding (RTB) advertising, estimating winning price is an important problem to estimate the bid cost and to determine the bid price of auctions in the bidding strategy on Demand-Side Platforms(DSPs). The state-of-the-art approaches utilize censored linear regression for winning price estimation by considering both successful and failure bid records. However, the traditional regression models are based on Gaussian distribution. We argue that the Gaussian distribution is not suitable for winning price estimation since the winning price is always positive and it is hard to link the physical meaning between Gaussian distribution and the winning price. In this paper, we propose a Gamma-based linear regression model which also considers both successful and failure bid records for the winning price estimation. To derive the parameters of our proposed model based on bid records, we adopt a two-step approach to simply derive the parameters. Furthermore, to reduce the number of iterations, we also provide several heuristic methods for initial parameter setting. The experimental results demonstrate that our approach is effective to estimate the winning price than the state-of-the-art approaches in two real datasets.