Predicting Winning Price in Real Time Bidding with Censored Data
1. Predicting Winning Price in Real
Time Bidding with Censored Data
Wush Wu#, Mi-Yen Yeh*, and Ming-Syan Chen#
#: Dept. of Electrical Engineering, National Taiwan University
*:Inst. of Information Science, Academia Sinica
4. Trading the Impression
● The sellers provide:
– Information of the
publishers
– Identification of the ad
viewer
● The buyers estimate:
– The value of the
impression
Bid Request:
● User Identity
● User IP
● URL
● Ad SlotVisibility
● Ad SlotSize
Advertisers
Publishers
Demand-Side
Platform (DSP)
Supply-Side
Platform (SSP)
Bid Response:
● Bidding Price
7. Winning Price
The highest bidding price from other competitors
● The winning price of purple: 200$
● The winning price of others: 250$
8. Our Goal: Predicting the Winning Price
● Predicting the winning price of future auctions given the
historical winning/losing bid information the buyer
observed
9. The importance of the Winning Price
● The winning price represents:
– the cost of the impression
– the value of the impression to the competitors
● The winning price helps the bidding strategy
● The winning price improves the estimation of the Click-
Through-Rate(CTR) and the Conversion Rate(CVR)
https://clientmanagementvn.files.wordpress.com/2012/09/competitor-analysis.jpg
10. Challenge of Predicting the Winning Price
● In second price auction, the winning price is unobserved if
the bid is lost.
● No previous work on predicting winning price on buyer
side
– Cui et al. modeled the winning price with the mixture-of-
log-normal distribution on various targeting attributes.
17. Mixture Model
● Censored regression model is closer to
unobserved data
● Linear regression model is closer to
observed data
18. Challenge of the Mixture Model
● We do not know whether the bid is winning
bids or losing bids
19. Winning Rate
● We use the estimated winning rate to classify
whether the bidding will be observed or
censored
– The winning rate is estimated by the
logistic regression
20. Mixture Model
● Learn the linear and censored regression models
● Learn the winning rate
● Combining these models to produce mixture model
22. Datasets
● iPinYou Real-Time Bidding Dataset
– Available at: http://data.computational-advertising.org/
– The codes for related experiments:
https://github.com/wush978/KDD2015wpp
● Bridgewell Inc., the major DSP in Taiwan
23. Preprocessing
● Use real winning bids only
● Set the bidding price to be x% of original bidding price
Original Bidding Price
Simulated Bidding Price
Original Winning Price, not changed
Original Bidding Price
Simulated Bidding Price
Original Winning Price, not changed
Simulated Losing Bids
Simulated Winning Bids
24. Questions
● (Q1) Different Winning Price Pattern
● (Q2) Censored regression model vs. linear regression
model
● (Q3) The Performance of the Mixture model
25. Inconsistent Pattern of Winning
Price (Q1)
● The avg. winning price is different on winning bids and
losing bids
Day Avg. WP on W Avg. WP on L
2013-06-06 52.46772 185.3269
2013-06-07 51.12051 186.9674
2013-06-08 58.48506 189.4200
2013-06-09 58.92701 188.2934
26. Inconsistent Pattern of Winning
Price (Q1)
● The performance of linear regression based on winning and
losing bids are different.
27. Censored Regression vs. Linear
Regression (Q2)
βlm is the linear regression
βclm is the censored regression
The MSE is evaluated on losing bids
28. Performance of the Mixture Model
(Q3)
βlm is the linear regression
βclm is the censored regression
βmix is the mixture model
The MSE is evaluated on losing
bids
- The mixture model usually
outperforms the linear
regression
- The mixture model is more
robust than the censored
regression
29. Conclusion
● We are the first to tackle the winning price prediction
problem from the buyer side
● Prediction performance is improved by taking the censored
information into account