Real Time Bidding : A Survey Pallavi Dr. Atul

 

Real Time Bidding : A Survey

Pallavi                                                                                        Dr. Atul Mishra
Deptt. of Computer
Engineering                                             
           Deptt. Of
Information Technology
YMCA University Of Science And Technology                                         
YMCA University Of Science And Technology
Haryana, INDIA,                                                                                  Haryana, INDIA,
[email protected]                                                                 [email protected]
 
     

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!


order now

Abstract—Real Time Bidding is the fastest growing area in the
online advertisement market. It has changed the way of advertisement from
traditional advertising to target audience buying. It is just similar to financial
market where the auction determines who gets to buy the impression and the bid
winner advertisement will be displayed on publishing site. So, it acts as an
interface between the buyer and seller of the ads.

In
this paper we will discuss the current RTB trends, various techniques of RTB and
the challenges associated with it.

 

Index Terms—Real-time bidding, Machine learning, bidding strategy, Impression,
Target Audience.

1.      
INTRODUCTION

 

Real-time bidding (RTB) is considered as the new
era of digital advertising 1. Digital advertising is one of the main means of
monetizing user data. RTB is a growing trend in today’s market which
compromises many different components requiring to effectively delivering
intelligent real time purchasing of advertisements

RTB has changed the way of advertising from
traditional advertising to target audience buying 2. In traditional
advertising, the cost of advertisement is previously negotiated between buyer
and seller but in RTB price is not previously fixed between advertiser and
publisher, it depends on the bid which is submitted by the advertisers.

Real time bidding used cookie based online big
data analysis for collecting and identifying the interests and aspects of
targeted audience and then shows ads that are suitable for them.

In this, the total numbers of impression clicked
by the user on any website are mapped with any advertiser using bidding
process. It allows advertiser to bid for slots available in web pages when user
launch the application. This advertisement is placed on user’s webpage only for
few milliseconds. This process is very complex as there are huge applications
available in the market. So, mapping the correct application with correct
advertiser on the basis of the content of the application is a difficult task.

RTB is becoming the key to target marketing
where it could optimize advertiser’s expectations drastically 3.

    Fig 1: RTB acts an
interface:

It acts as an interface between advertisers and
user’s of the websites and makes the match that what advertisement is shown to
user according to his interest so that it can generate maximum number of
impressions.

RTB is the real time
process which gives only less than 120 ms for deciding the price of bid 3.
So, it is not useful for that practices which takes more time in decision
making.

 

2.       BUSINESS MODEL OF REAL TIME BIDDING

 

Under this section, we will give detailed overview
about key roles and business flow of RTB process.

 

2.1     KEY ROLES OF REAL TIME
BIDDING MARKETS

 

In
this section we will give detailed explanation about roles of various RTB
entities such as advertiser, publisher, DSP, DMP and Ad exchange.

Fig 2: RTB ecosystem Key role in RTB markets:

The
role of various entities in RTB markets are as follows:

Advertiser:
is the buyer of ad slots on which their ad will be displayed. They participate
in ad auction and bid according to their budget, market goals etc. the
advertiser who bids highest amount wins the ad impression.

Demand
side platform (DSP): it helps advertisers in buying the appropriate ads from ad
exchange in very efficient way. It provides match on the basis of big data
analysis, audience targeting etc.

Ad
Exchange: it acts as an interface between user and DSP. It displays ads
according to the interest of the user.

Publisher:
is the owner of the website searched by the user. When a user launched the ad
of advertisers who wins the ad auction will displayed on the ad slots.

Data
management platform (DMP): is a data warehouse and store, merge and label user
information which is useful to advertisers and DSP. This information is
collected by cookie synchronization method.

 In cookie sync method, when a user opens a
webpage and goes without closing it then a third party tracker tracks the
user’s activities and stores the information in DMP.

 

2.2    BUSINESS FLOW OF RTB PROCESS

 

As seen in fig 2 the detailed flow of RTB process is
given as follows:

1.      
Firstly any user on internet visits the webpage and this webpage is
owned by a publisher.

2.      
If on this webpage ad slots are available then publisher send this
information to ad exchange with user’s information, number of slots and
demanded price.

3.      
Ad exchange then forwards it to all DSP’s.

4.      
Each DSP parses the information given by ADX. DSP ask DMP for more
information about the user such as location, age, sex, shopping interest etc.
after finding information it places an auction among all the advertisers which
are associated with DSP. The information about the winner of the auction is
provided to AdX.

5.      
ADX is connected to several DSPs and receive winner from each DSP. Then
it again starts an auction among all winners of DSPs, if the bid of advertisers
is less than the publisher’s demanded price then it terminates the process else
it notifies the DSP to the winner. Then it plays the advertisement of winner in
the ad slot of the user.

 

3.      
Literature Survey

 

Various RTB techniques have been proposed for
bidding strategy. The different techniques are described below:

 

Myerson et al 4 proposed an optimal mechanism.
In optimal mechanism the bidder who bids highest price wins the auction but
gives only second highest price. There are two stages in auction process. First
is DSP advertisement and second are advertisement exchange/ DSP stage. There is
no incentive given to the DSP who truthfully submits the bids to all of its
advertisers. This leads to decrease in adv. exchange revenue.

 

Mansour et al 5 prove that the optimal second
price forces all the advertisers to truthfully submit their price as we are not
making highest winning price available to advertisers. So, there is less
information available to advertisers about winning price. They bid their higher
price in order to win the bid. So, in this way loss of adv. Exchange will be reduced.

 

Researcher introduced hybrid BIN TAC mechanism
for increasing the revenue of ad exchange 6. In this impressions are
auctioned by win using buy it now price. If only one advertiser can give that
price, then he/she wins the auction. If many advertisers want to purchase that
impression, than a second auction is held between them. If there is no
advertiser who can afford that price, then a take a chance auction is held
between top L advertisers. The impression is randomly given to any advertiser
at L+1 st price.

 

Yuan et al 7, 16 proposed time dependent model
for evaluating the ad performance. It observed the ad impression, click and
conversion rate for evaluation.

 

Rogers et al. 8, 18 have suggested a
probabilistic model by considering the users behavior and advertisers. This
model focusses on bidding strategy and find out most reasonable bid value for
each auction and also determines how we can achieve maximum number of
impressions.

 

Hegeman et al. 9 have suggested the standard
for the bidding strategy. It considers the past value of impressions, total
allocated budget, presence of social functionality, chances of advertisement
selection and total available budget etc. It considers the past results in
which DSP wins the auction and estimate the bid price of its current auction.
In some other research bidding strategy was made by best using values of budget
and price of the bid.

 

Chakraborty et al. 10, 17 suggested the
optimization framework. In this framework, they combined online algorithms and
stochastic models. They proposed Bayesian decision framework with constraint of
type of bandwidth. In Bayesian framework we display advertisements according to
the content of the web page which is searched by the user.

 

 Li et al
11 tries to forecast bid value by taking an impression at lowest cost. They
use logistic regression model for forecasting winning price and win rate model
for forecasting win rate. After combining both the models, they tried to
estimate bidding strategy.

 

R. Schapire el at 12, 13, and 14 has proposed
one high level approach which is based on machine learning. This approach uses
data of ad exchangers as training data for win rate estimation model. This data
can be either negative or positive. By using this data, a toolkit for
generating bid price for further advertisements is made.

 

Chen et al. 15 proposed impression level bid
price issue as a constraint optimization problem as there are many constraints
such as limit of budget, availability of inventory for maximizing revenue.
Their bidding algorithm is based on linear programming this algorithm proposed
detailed impression valuation and consider the constraint for suggesting real
time bid value.

This is an online
algorithm which guarantees to give provide same level of knowledge as offline
optimization.

 

Claudia et al. 16
purposed a bid optimization technique which is based on supervised learning and
second price auction theory. This approach gives effective results in matching
the impressions to audience. This approach is used in Media6Degree platforms.

4.      
RTB AUCTION MECHANISM

In online marketing buyers and sellers do not
know the exact price for an ad impression in market. In such situation, auction
is held among buyers of ad impression. Auction is fair, simple, widely used and
transparent way in which buyers and sellers are agreed on a price quickly in
order to increase their sales outcome.

In auction we offered the item which we want to
sell to bids and give that item to that buyer who pays the highest price.
Different auction schemes are as follows:

The second price auction model: in this, many
advertisers bids for an ad impression if anyone from them wins the bid then
he/she should give only second highest bid price.

For example there are 4 advertisers (P, Q, R, S)
who are bidding for same ad impression with bidding amount 10$, 8$, 9$, 16$
respectively. The fourth advertiser will win the auction and give only 10 $ as
bidding price. The second bidding price becomes the sealed price or market
price.

Winning probability: In the above model we
calculate the market price per ad impression basis but in this we calculate the
market price distribution per campaign basis.

In this we use winning
probability for click through rate estimation.

      Bid landscape forecasting: in this, we
estimate the performance of different ad groups in different ad scenario. It is
basically used for bid optimization process. Suppose your ad is running from
few weeks in a certain ad slot but it is becoming more costly week by week.

    Now you want to see the effect on your ad
impression by reducing bid price a little. You want to see effect on number of
clicks, cost etc.

    You also want to see that you are getting
enough impression after reducing your price as well or not then bid landscape
forecasting helps you. It tells us that by seeing past responses of ads.

5.      
BIDDING STRATEGY

A bidding strategy refers to the selecting the
accurate bid price for an ad impression so that the probability of winning an
auction will be increased. The design of optimal bidding strategy is most
difficult and important work in RTB.

 

 

 

Fig 3: bidding strategy as function

Bidding strategy is an optimization function. We
can assume biding strategy as a function which takes bid request (user
information, ad page on which ad is displayed, context) as input and provides
the bid price as output. Various types of bidding strategies are as follows:

Quantitative biding in RTB: in performance
driven environment each advertisement is quantified by utility and cost.

Utility is the chances of a user clicked on ad
or not. Cost is cost spent in winning the ad impression.

Strategy for profit
maximization: the main motive of any business is to earn profit. So we want a
bidding strategy which makes maximum profit from daily business operations.

    Profit is calculated as difference between revenue
and cost. If the difference is more than the profit is also more and it’s vice
versa. But only if our bidding price is higher than our chances of winning an
auction is increased.

    Strategy for revenue maximization: In this,
we try to maximize revenue. An advertiser places a bid according to first price
auction mechanism which means it maximize its bidding price in such a manner
that it cannot lose money.

    Strategy for profit and revenue
maximization: In this, we try to maximize both profit as well as revenue. So we
mix both of the above function together and use a weight alpha to control the
strategies. When we placed a bid then we can adjust alpha value for controlling
the importance of profit vs. revenue. This strategy can change value of alpha
in real time depends on performance of bidding system.

 

6.       PREDICTION OF USER’S RESPONSE

Learning and predicting user’s response is most
difficult in online marketing. In this, we have to learn how to estimate user
response such as clicks conversion, readings etc.

We have to predict user response accurately so
that profit of the advertiser can be maximized. We can use CTR and CVR
prediction for performance driven marketing. Various approaches for predicting
user response are as follows:

 

 

 

 

 

 

 

 

 

 

 

 

Table 1: Comparison of different Types of User
Response Prediction Model

Type
of model

Purposed
By

Method

Usage

Issues

Bayesian probit regression

It is purposed by Grapetet et
al., in 2010.
It is also called a Bayesian learning
model

It is a linear model.
        

 

It can easily implement with high
efficiency.

They learn with very few features pattern
and they are inefficient for catching high order patterns.
Solution: This issue is solved by non
linear model such as factorization machines

Factorization machines

It is purposed by Rendle in 2010.

It directly explores features and
interactions among them and mapped these features and relationship in low
dimensional space.
 

It is very useful in collaborative
filtering as it is an extension matrix factorization.
 

CTR estimation for mobile
advertisement is a problem.
Solution: This problem is solved
by extending FM to HIFM (hierarchical importance aware factorization
machine).
 

Decision tree

It is purposed by Breiman et al., in
1984.

It is a non linear supervised
learning method.
 

It is used for CTR estimation.
 

Addition and deletion of specific
feature is a problem.
Due to small change, tree will
break in parts.
Solution: its solution is
ensemble learning.
 
 

Ensemble learning

It is purposed by Friedman et al., in
2001.
 

In this, many decision trees are
combined together for solving problem of splitting and avoiding over fitting
in data.
It uses two techniques for that
which are as follows: bagging and boosting.
 

It is useful for decreasing variance and
over fitting of data.
 

It helps with increasing the volume of
training set but predictive force of model will not improve.
It requires less correlation between
trees.
 

Bagging

It is purposed by Breiman in 1996

It is more powerful and useful than
boosting.
It is also known as bootstrap
aggregating.
In this we will take bootstrap samples
and train each sample. After training we will estimate the average of
prediction of these different algorithms.
 

It is useful for decreasing variance and
over fitting of data.
 

It helps with increasing the
volume of training set but predictive force of model will not improve.
It requires less correlation
between trees.
 

Boosting:

It is purposed by Breiman in 1996.

In this the output of many weak
predictions are combined so that a strong prediction is produced.
It works iterative manner.
 

It also gives higher stable model by
reducing variance:

It does not help in protecting over
fitting.

User Look alike

It is purposed by Zhang et al., in 2016,
Mangalanpalli et al. (2011).

RTB has the ability to make user
profiles and detects user interest via tracking user’s behavior by history,
clicks and conversions.
 

It provides high targeting
accuracy which results more number of conversions by the user.
 

In this building of user interest segment
is performed independently which results it has less attention towards
prediction of ad response.

Transfer of learning process from
browsing of web to ad clicks

It is purposed by Pan and Yang in 2010.

Here
we learn by transferring the knowledge from other related tasks. It is
related to multi-task learning.
In
RTB training data is transferred by two resources.
One
is user behavior and other is ad response by user. User behavior estimation
is task of collaborative filtering and ad response prediction is task of
CTR.                                 
 

It solves the problem of classification,
regression and collaborative filtering.
It is useful in that situation where
training data for learning process is very expensive and quickly outdated.
 

The amount of observations are very large
it is quite promising to deal with this large amount of data.

 

 

 

7.      
ATTRIBUTION MODEL

 

The user’s interaction on a particular ad is
tracked by the number of clicks. The particular ad item is purchased by user or
not is determined by the touch points. So credit of conversion is allocated
over touch points. We have to make good conversion attribution in online
marketing. Two different types of models are as follows:

Heuristic model: In this model, various rules
are created by human by suing their past business experiences. Example of such
type of heuristic model is Google Analytics.

Data driven probabilistic model: There are two
types of data driven probabilistic model are as follows:

Bagged logistic regression: It tries to predict
user’s behavior in terms that he/she is converting ad touch event or not.  It does not consider repeated touch of same
user. The user’s ad touch event is given as input to the regression process and
output is a binary value which determines user’s response in yes or no.

Simple probabilistic model: it uses both types
of conditional probabilities for credit allocation.

8.      
 conclusion

Real Time Bidding helps
to simplify the bidding process and help marketers to take their revenues to
new heights. In this paper, we have briefly introduced RTB marketing model and summarize
the related work undergoing in this field. It is connecting many platforms such
as desktop, mobile web,

and social media. But
here are many open issues in this field. In this, we have dealt with various
auction mechanisms and bidding strategies. We also read about comparisons
between various user response models. There are many issues in RTB, one such
issue is the tradeoff between the innovations in procedures of advertisement
and security of the user’s personal data.

 

 

 

 

REFERENCES

 

1    
Real Time Bidding in Online Digital
Advertisement Shalinda Adikari and Kaushik Dutta National
University of Singapore, Singapore, Singapore

2    
A Survey on Real Time Bidding Advertising Yong
Yuan1,2, Feiyue Wang1,2, Juanjuan Li1, Rui Qin1,3 The State Key
Laboratory of Management and Control for Complex Systems, Institute of
Automation, Chinese Academy of Sciences, Beijing, China  Qingdao Academy of
Tntelligent Industries, Qingdao, China Beijing Engineering Research Center of
Tntelligent Systems and Technology, Institute of Automation, Chinese Academy of
Sciences, Beijing, China

3    
Real Time Bidding in Online Digital
Advertisement Shalinda Adikari and Kaushik Dutta National
University of Singapore, Singapore, Singapore

4    
Muthukrishnan, S., Ad exchanges: Research
issues. Internetandnetwork economics, 2009: p. 1-12.

5    
Mansour, Y, S. Muthukrishnan, and N. Nisan,
Doubleclick Ad Exchange
Auction. Arxiv preprint arXiv: 1204.0535,2012.

6    
L.Elisa c., et al., Buy-it-now or
Take-a-chance: A New PricingMechanism for Online Advertising. Working Paper,
2012.

7    
Yuan, S., J. Wang, and X. Zhao. Real-time
bidding for online advertising: Measurement and analysis. in 7th Internationa
Workshop on Data Mining for Online Advertising,A DKDD 2013 -Held in Conjunction
with SIGKDD 2013,A ugnst II,2 013 –August 11, 2013. 2013. Chicago, IL, United
states: Association for Computing Machinery.

8    
Rogers, A., David, E., Payne, T.R., Jennings, N.R.: An advanced bidding
agent for advertisement selection on public displays. In: Proceedings of the
6th International Joint Conference on Autonomous Agents and Multiagent Systems,
p. 51. ACM (2007)

9    
Hegeman, J., Yan, R., Badros, G.J.: Budget-based advertisement bidding,
US Patent US20130124308 A1 (2011)

10  Chakraborty,
T., et aI., Selective call out and real time bidding.Internet and Network
Economics, 2010: p. 145-157.

11  Li, X., Guan,
D.: Programmatic Buying Bidding Strategies with Win Rate and Winning Price
Estimation in Real Time Mobile Advertising. In: Tseng, V.S., Ho, T.B., Zhou,
Z.-H.,Chen, A.L.P., Kao, H.-Y. (eds.) PAKDD 2014, Part I. LNCS, vol. 8443, pp.
447–460. Springer, Heidelberg (2014)

12  Schapire, R.,
Stone, P., McAllester, D., Littman, M., Csirik, J.: Modeling auction price
uncertainty using boosting-based conditional density estimation. In: ICML 2002
(2002)

13  Wellman, M.,
Reeves, D., Lochner, K., Vorobeychik, Y.: Price prediction in a trading agent
competition. Journal of Artificial Intelligence Research 21, 19–36 (2004)

14  Putchala, R.,
Morris, V., Kazhanchi, R., Raman, L., Shekhar, S.: kavayaH: A trading agent developed for
TAC-02. Tech. Rep., Oracle India (2002)

15  Chen,
Y, et al. Real-time bidding algorithms for performance-based display
ad allocation. 2011. ACM.

16  Claudia, P et
aI,., Bid Optimization and Inventory Scoring in Targeted Online Advertising,
in The 18thACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2012: Beijing, China.

17  Chakraborty,
T., Even-Dar, E., Guha, S., Mansour, Y., Muthukrishnan, S.: Selective callout
and real time bidding. In: Saberi, A. (ed.) WINE 2010. LNCS, vol. 6484, pp.
145–157. Springer, Heidelberg (2010)

18  Ghosh, A.,
McAfee, P., Papineni, K., Vassilvitskii, S.: Bidding for representative allocations
for display advertising. In: Leonardi, S. (ed.) WINE 2009. LNCS, vol. 5929, pp.
208–219. Springer, Heidelberg (2009)

 

 

 

 

x

Hi!
I'm Angelica!

Would you like to get a custom essay? How about receiving a customized one?

Check it out