1. Predicting Winning Price In Real Time Bidding With Censored Data Tejaswini Veena Sambamurthy Weicong Chen

2. Topics Covered Introduction Problem Statement Specification Methodology  Linear Regression Model  Censored Regression Model  Mixture Model Datasets Experiments  Settings  Winning Price Patterns  Evaluations Related Work Reviews

3. Introduction

4. What is RTB?  Real-time bidding (RTB) is a relatively new method of selling and buying online display advertising in real time one ad impression at a time. What is SSP?  Publishers manage and sell their inventories of ad impressions via the Supply-Side Platform (SSP) What is DSP?  Advertisers can bid and manage their ads across multiple inventory sources via the Demand-Side-Platform (DSP)

5. Introduction Censoring  The winning price is only observable for the DSP who wins the bid. For the lost DSP they can only observe the lower bound, which is their own bidding price.  Censoring also occurs because of the use of soft floor price  The publishers can set a soft floor price and the hard floor price for each ad impression. Hard floor price is the minimum bidding price to even have a chance to win. The soft floor price is higher than the hard floor price and its usage can be understood from the flow chart in the next slide.

6. Introduction Mechanism of modern RTB

7. Problem Statement The goal of this paper is to learn the winning price, in the aspect of a DSP with historical RTB bids and to deal with the problem of various winning price censoring at the same time. We aim to propose a machine-learning based model that can help a DSP to predict the winning price and thus decide the bidding price.

8. Specifications How to solve these difficulties?  A linear regression model based only on the observed bids  A censored regression model to consider the censored data.  A mixture model which combines the two models, weighted by the winning rate of the DSP. Experiments were conducted on real RTB data sets.  Result shows the mixture model in general prominently outperforms linear regression in terms of prediction accuracy

9. Math Terminologies I Regression: In statistical modeling, regression analysis is a statistical process for estimating the relationships among variables. Regression analysis helps one understand how the typical value of the dependent variable changes when any one of the independent variables is varied, while the other independent variables are held fixed. Likelihood function: A likelihood function is a function of the parameters of a statistical model given outcome. Probability function: describing a function of the outcome given a fixed parameter value

10. Methodology: Modeling winning price

12. How to measuring how good the linear regression model can approximate? Negative log-likelihood function

13. Methodology: Modeling Censored Data So modeling censored data is necessary!

14. Methodology: Modeling Censored Data It measures how well the linear regression model approximates the censored data.

15. Math Terminologies II

16. Methodology: Modeling Censored Data

17. Methodology-Mixture model

20. Dataset

21. Basic Statistics of the iPinYou Dataset Estimate Winning Rate Winning Rate Area Under Curve Average winning price Winning bid Censored bid

22. Experiment Settings Since both the DSPs can only provide the winning prices of their own won bids, simulation of the bidding results of both winning and losing bids is processed as follow: The new bidding price is compared to the original winning price to produce the simulated winning and losing bid. The proposed mixture model requires the winning rate as input which is predicted by the simulated bidding result. CTR (click-through-rate) is an important feature for predicting winning rate. Predict CTR and the winning rate in an online fashion to simulate the online bidding environment. For each historical bid, divide the bidding price originally offered by each DSP by a factor as the new bidding price (default: 2).

23. The difference of Winning Price Patterns Question: Are the patterns of the winning price on the observed data and those on the censored data different? The average winning prices on losing bids are usually higher than the average winning prices on winning bids.

24. The difference of Winning Price Patterns

26. Results show that the winning price patterns on the winning and losing bids are different, otherwise the result of the two estimators should be similar. To ameliorate this effect, interaction between the estimated winning rate and the features used to train the model is considered. The model is trained by considering both the original features and the features after interaction with the estimated winning rate. The different patterns of winning and losing bids can thus be differentiated.

27. Does the censored regression model have better performance on predicting winning price compared to the linear regression model trained only from the historical winning bids?

29. Evaluation on Mixture Model

32. Related Work Zhang et al. Used bidding price as the upper bound of the cost as the winning price. Ghosh et al. Considered both the fully and partially observed winning price information. Since their main goal was to consider budget constraints, they simply assumed the winning price was drawn I.i.d from a CDF. Other studies are on the seller side (SSP)

33. Thank you! Questions?