In this assignment, you are required to work on an online

COMPGW02/M041 Web Economics Coursework
Lecturer: Prof. Jun Wang
jun.wang@cs.ucl.ac.uk
TAs: Yean-Hoon Ong, Zheng Tian, Minne Li and Ying Wen
{yean-hoon.ong, zheng.tian.11, minne.li.17, ying.wen.15}@ucl.ac.uk
Submission Deadline: 23:55 06/04/2018
This coursework project consists of two reports and is worth 30% of your overall score
1. SUMMARY
In this assignment, you are required to work on an online
advertising problem. You will help advertisers to form a
bidding strategy in order to place their ads online in a realtime bidding system. You are required to train a bidding
strategy based on a provided advertising impression training
set. This project aims to help you understand some basic
concepts and write a computer program in real-time bidding
based display advertising. As you will be evaluated both as a
group as well as individually, part of the assignment is to train
a model of your choice independently. The performance of the
model trained by your team, which is either a combination
of the individually developed models or the best performing
individually-developed model, will be (mainly) evaluated on
the Click-through Rate achieved on a provided test set. For
you to properly evaluate the performance of each of your
models before that, a benchmark click-through rate on the
validation set will be provided. Before the final submission,
you are also given the opportunity to hand in the preliminary
result of your team’s model on the test set, which allows you
to compare the performance to that of your peers. Should you
have any question about the assignment, feel free to contact
the TAs: Yean-Hoon Ong (email), Zheng Tian (email), Minne
Li (email), and Ying Wen (email).
2. RULES
You are expected to work in a team of 3 people, and you
can sign in the team through this google sheet, by the end
of this week (the 9th Feb 2018). You will write a joint group
report as the group, and also an individual report on your
own.
2.1 Submissions (In Moodle)
• Preliminary submission: 23:55 16/03/2018. Please upload your submission through Moodle as a zip file, and
rename the file as that: Group xx.zip (e.g., Group 10.zip).
For the preliminary submission, only include a bid
price generated using your team’s model applied to the
testing set in the zip file. Note that the preliminary
submission is for the practice purpose, making sure the
final submission would be correct. We are working on
an alternative online submission for a real-time update.
stay tuned.
• Final submission: 23:55 06/04/2018. Both the Group
Report and Individual Report will be submitted together. Please upload your submission through Moodle
as a zip file, and rename the file as that: Group xx.zip
(e.g., Group 10.zip). For the final submission, the file
should include the reports (both the group and individuals) in PDF format (including the names of all
members of your team, the reports should be named as
group xx report.pdf and your name report.pdf), and
your team’s predictions for the provided test set. Please
upload your code to GitHub and provide a link to it in
the reports (again for both reports).
2.2 The Group and Individual Reports
Individually, you are required to submit an Individual
Report (one per person). At the same time, as a group, you
are required to submit one Group Report (one per group).
Each of them will have 15 marks, and in total there are 30
marks. The rest 70 marks will come from the written exam
in May.
The two reports will have different sets of questions, respectively, as given in Section 3.2. The reports shall summarise
what you (or your group) did and how you addressed the
posed questions as described in the problem description.
You will use the ACM Proceedings style LATEX files for
writing your reports. You can find out more about these
styles as well as LATEX templates on the ACM website.
Note that to make your report looks better, you could simply
add the following lines before \begin{document}} to remove
the copyright notice:
\makeatletter
\def\@copyrightspace{\relax}
\makeatother
For the Individual Report, your submission should be 3-5
pages, with maximum 1 additional page for references. For
the Group Report, your submission should be 5-7 pages, with
maximum 1 additional page for references. Images and tables
also count against the maximum. Rules written here trump
all the non-formatting instructions in the ACM style files.
Your report should address all the problems in Section 3.2.
2.3 Q&A and Office Hours of TAs
The office hours below are for this project only. For the
office hours of the module itself, please refer to the Moodle
page.
• Time: 12:30-13:30 20/02/2018 06/03/2018 and 20/03/2018.
• Location: Hub Level 1, 66-72 Gower Street.
3. THE PROJECT
In display and mobile advertising, the most significant
evolution in recent years is the employment of the so-called

Real-Time Bidding (RTB) mechanism to buy and sell ads. RTB essentially facilitates buying an individual ad impres sion in real time, automatically triggered by a user’s visit.
Although other types of auctions, such as the first price
auction, are also popular, RTB exchanges typically employ the second price auction model. In this type of auction, the bidder winning the auction pays the price of the second highest offer. A summary of the process is provided in Figure	Figure 1: A brief illustration of the interactions be tween user, ad exchange and the advertiser’s DSP
1. In this assignment you will be training your own bidding
bidding agent.

 

strategy models on a given impression dataset, as described

in detail next.
3.1 Data Description
The dataset includes a training set, a validation set, and
a testing set. The dataset for the task can be downloaded
from https://goo.gl/aqGmDX.
This data comes in CSV format, the first line in the file
containing the header formatted as described in Table 1.
As the testing set is used for final evaluation purposes, it
does not contain the three fields: ‘bidprice’, ‘payprice’ and
‘click’. Note that that all number related to money (e.g., bid
price, paying price and floor price) is Cost Per Mille (CPM),
corresponding to the commonly adopted CPM pricing model.
It uses the currency of CNY and the unit is Chinese fen.

For instance the payprice in Table 1 means the cost is 322
Chinese fen per 1000 impressions. When someone won this impression, they had to pay 0.322 Chinese fen just for it. bid price. Explain how you can experimental find an optimal constant value from the training set. Provide
You can develop your model on the training data and use
the validation set to compare the different models and correct an analysis of your finding.
for overfitting. 3.2 Problem Description	• Random bidding (Rand). Randomly choose a bid value in a given range. The parameter is the upper bound
In this assignment, you are required to train a bidding your experiment results of finding the optimal bounds.
strategy based on the impressions in the training set. This ob
jective has been divided into the following five sub-problems, and you are required to solve and report these problems in 3.2.3 Problem 3: Linear Bidding Strategy
either the group or individual report, as indicated below. The concept of a bidding strategy has been visualized in	– Group report, 5/30 marks
and lower bound of the random bidding range. Give

 

Figure 2 and can also be found in the lecture notes. 3.2.1 Problem 1: Data Exploration

Apply CTR estimation to create a linear bidding strat egy. The bid value is linearly proportional to the pCTR (predicted CTR). The formula can be generally written as bid = base bid×pCTR=avgCTR, where the tuning parameter

 

– Individual report, 5/30 marks

base bid is the bid price for the average CTR cases. Opti mise the base bid and the CTR estimation and evaluate the

You should do some exploratory data analysis and show
your understanding of this dataset. You can start from basic
statistical information, such as num Imps, num Clicks, Cost,
CTR, avg CPM, eCPC. You can also include further analysis
on user feedback, bidding etc. A good example can be found
in the paper [1].
3.2.2 Problem 2: Basic Bidding Strategies
– Group report, 5/30 marks
Evaluate the following strategies on the validation set and
discuss your results. Use the performance metrics as defined
in Section 3.3.
• Constant bidding (Const). Bid a constant value for all
the bid requests. The parameter is the specific constant
1.2. Real-Time Bidding based Display Advertising 25
Demand-Side
Platform
1+. User Information
Advertiser
RTB
Ad
Exchange
Data
Management
Platform
1. Bid Request
(user, page, context data)
(demographics, browsing, interest segments)
(ad, bid price)
2. Bid Response
3. Auction
0. Ad Request
5. Ad
6. User Feedback
Page
User
(with tracking)
4. Win Notice
(charged price)
(click, conversion)
Ad
Figure 1.3: A brief illustration of the interactions between user, ad exchange and the advertiser’s DSP bidding agent. The solid-line arrows represent real-time interactions,
while the dashed-line arrows represent the interactions that are not necessarily in
real-time. The step (1+) is optional, i.e., the DSP may use their own user segmentation information instead of calling the third-party DMP for this user information.
Besides, this figure is mainly from a DSP’s perspective, thus some components,
such as supply-side platform (SSP) and ad server are omitted for simplicity.
web pages, streaming videos and mobile apps), each ad placement will trigger a call
for an ad (ad request) to the ad exchange. (1) The ad exchange sends the bid requests,
for this particular ad impression, to each advertiser’s DSP bidding agent, along with the
other available information such as the user cookie and context information. (2) With
the information of the bid request and each of its qualified ads,1 the bidding agent calculates a bid price. Then the bid response (a pair consisting of the ad and the bid price)
is sent back to the exchange to take part in the auction. (3) Having received the bid
responses from the advertisers within a predefined time window, the ad exchange hosts
an auction and picks the ad with the highest bid as the auction winner. (4) Then the
winner is notified of this result and the price which will be charged by the ad exchange.
(5) Finally, the winner’s ad will be shown to the visitor along with the regular content
of the publisher’s site. It is commonly known that a long page-loading time will greatly
reduce the user’s’ experience, in terms of quality [16]. Thus, advertiser bidding agents
are usually required to return a bid in a very short time frame (e.g., 100ms). (6) The
user’s feedback (e.g., click and conversion) on the displayed ad is tracked and finally
sent back to the winning advertiser. A more detailed discussion concerning the whole
RTB eco-system is given in [14].
From the above interaction processes, it is obvious that, for DSPs, the most signif-
1The qualification of each ad for the bid request is based on ad size matching, the campaign’s target
rules, etc.
54 5. BIDDING STRATEGIES
Bidding Strategy
Bid Request
Bidding Function Bid Price
(user, ad, page, context)
INPUT
OUTPUT
Data Preprocessing
Utility
Estimation
Cost
Estimation
Figure 5.2: The quantitative bidding: the logic of the bidding function only
depends on two (sets of) factors, i.e., the estimated utility and the cost of
the ad display opportunity.
For example, a sneakers advertiser would like to bid high on users with ages
between 15 and 30; this is motivated by the fact that users in such a segment
are more likely to be converted to purchase the sneakers after seeing the ads.
This is quantified as a higher conversion rate. This is analogous with a
high frequency trading strategy in a stock/option market where the trading
action is wholly based on the quantified risks and the returns for each asset,
regardless of the specific asset attributes or fundamentals [Durbin, 2010].
Using the estimated utility and cost, the optimal bidding function to
maximise the specific KPI under the target campaign budget and auction
volume constraints can be derived.
The research frontier and implementations of the utility estimation module and cost estimation module in Figure 5.2 have been well investigated in
Chapter 4 and Section 3.3, respectively. In the later sections of this chapter,
we will discuss various bidding strategies and their connections.
5.3 Single-campaign bid optimisation
With the modules of CTR/CVR estimation (Chapter 4) and bid landscape
forecasting (Section 3.3) ready, bid optimisation techniques seek an optimal
bidding strategy to maximise a certain objective (with possible constraints).
In this section, we will discuss the derivation of different bidding strategies under their presumed settings.
5.3.1 Notations and preliminaries
Define the bidding function as b(r) which returns the bid price given the
estimated click-through rate (CTR) r of an ad impression.
Winning probability. From an advertiser’s perspective, given the market
Figure 2: A bidding strategy can be abstracted as
a function mapping from the given bid request (in
a high dimensional feature space) to a bid price (a
non-negative real or integer number).
performance of your model using the metrics as defined in
Section 3.3. Discuss and compare your results.
3.2.4 Problem 4: Your Best Bidding Strategy
– Individual report, 10/30 marks
As the linear model may not be the best model, you are
asked to experiment further in order to find a more optimal
model. The model you developed can be tested over the
validation set. and you should explain your approach and
discuss your results in the individual report. This does
include, but is not limited to, reporting the performance
metrics as defined in Section 3.3 of your solution and the
parameter turning results on the validation set.
Table 1: Fields in dataset
Field Example Supplement
click 1 1 if clicked, 0 if
not.
weekday 1
hour 12
bidid fdfe…b8b21
userid u Vh1OPkFv3q5CFdR
useragent windows ie
IP 180.107.112.*
region 80
city 85
adexchange 2
domain trq…Mi
url d48a…4efeb
urlid as3d…34frg
slotid 2147…813
slotwidth 300
slotheight 250
slotvisibility SecondView
slotformat Fixed
slotprice 0
creative hd2v…jhs72
bidprice 399
payprice 322 Paid price after
win the bidding.
keypage sasd…47hsd
advertiser 2345
usertag 123,5678,3456 Contains multivalues, ‘,’ as segmentation.
Note that your bidding strategy described here may be part
of the more complex model (either “cleverly” combined one
or further developed one based on multiagent reinforcement
learning) for the group, see problem 5 below, which will
be tested over the test dataset. Therefore, it is advised,
although not mandatory, that you try different models than
your other team members in this section.
Some directions: A non-linear bidding strategy (e.g. ORTB)
[2], a lift-based bidding [3] or using other CTR estimators.
A collection of research papers can be found here.
3.2.5 Problem 5: A Further Developed Bidding Strategy (Submission)
– Group report, 5/30 marks
As a group, you can either 1) compare the individual
bidding strategies developed in problem 4 and design a combined strategy. Your report should include the discussion
of the way you combine them and the performance of your
combined model on the validation set. or 2) develop a (multiagent) reinforcement learning (RL) approach for the bidding
(you might want to make use of the basic bidding algorithms
developed in problem 4 to start with). Hint 1: RL can be
used to deal with budget constraints [4]. Hint 2: multiagent
RL might be useful for taking into the competition among
different group submissions.
Next, pick up the best bid strategy, and this can be the
combined model or the (multiagent) RL based solution and
make your predictions on the given test set. (There are also
the predictions that you are allowed to hand in preliminary.)
3.3 Evaluation
Alongside the final report, you are required to submit the
bid prices for the testing set, generated using your group’s
best performing model from problem 5. Performance will be
evaluated based on the following metrics, within a limited
budget of 6; 250 CNY fen.
• Clicks (Num. of Clicks from Winning Impressions) the
main metric to rank the submissions
• Click-Through Rate (Num. of Clicks over Winning
Impressions)
• Spend (Total Money Paid) spend all the budget to
be a valid submission
• Average CPM (Cost Per Mille)
• Average CPC (Cost Per Click)
The bidding prices on the testing set (not the training
or validation set!), should be saved as a CSV file named
‘testing bidding price.csv’ with following format:
bidid,bidprice
xxxxxxxxxx,13.0
yyyyyyyyyy,23.0
….
An example submission file will be provided with the
dataset, please strictly follow the format requirements. During the evaluation, we (our script) will go through the file and
only consider the bids until the accumulated cost is higher
than or equal to the allowed budget 6; 250 CNY fen.
We would expect you to produce two final submissions
(therefore files) to account for two different ways of calculating the winning and the actual paid price. Winning
criterion #1: the winning is determined if bid ≥ payprice
and the actual paid price is the payprice from the data.
Winning criterion #2: the winning is determined if bid ≥
payprice&otherSubmittedBids, and pay the highest among
payprice&otherSubmittedBids. Note that the latter setting
is still the second price auction, but accounts for the competition among the bids coming from different groups as specified
by Problem 5 option 2. If you don’t choose option 2, we still
require you to submit the result for Winning criterion #2,
but only make it the same as the one for Winning criterion
#1.
3.4 Format of your reports
Your reports (both the individual report and the group
report) should consist of the following sections: 1) the introduction section briefly explains the problem context and
what you intend to do and summarises your results; 2) the
related work section briefly summarises the related work of
the specific problems you have found from the research literature. You could use https://ift.tt/1QIwUwI or make
use of the list here to obtain the related research papers; 3)
the approach and result section gives your answers to each
of the specific questions and provide the evaluation results
and your discussions. 4) the conclusion section concludes
your report and point out the potential direction to improve
your report (e.g., if you have time, you will do…). For the
individual report, in the conclusion section, create a paragraph to explain your role in the group project and also, if
possible, comment on other members’ contribution to the
group. We will gather the information to understand how
the collaboration has been done in the group.
4. REFERENCES
[1] Weinan Zhang, Shuai Yuan, Jun Wang, and Xuehua
Shen. Real-time bidding benchmarking with ipinyou
dataset. arXiv preprint arXiv:1407.7073, 2014.
[2] Weinan Zhang, Shuai Yuan, and Jun Wang. Optimal
real-time bidding for display advertising. In Proceedings
of the 20th ACM SIGKDD international conference on
Knowledge discovery and data mining, pages 1077{1086.
ACM, 2014.
[3] Jian Xu, Xuhui Shao, Jianjie Ma, Kuang-chih Lee, Hang
Qi, and Quan Lu. Lift-based bidding in ad selection. In
AAAI, pages 651{657, 2016.
[4] Han Cai, Kan Ren, Weinan Zhang, Kleanthis Malialis,
Jun Wang, Yong Yu, and Defeng Guo. Real-time bidding
by reinforcement learning in display advertising. In
Proceedings of the Tenth ACM International Conference
on Web Search and Data Mining, pages 661{670. ACM,
2017.

The post In this assignment, you are required to work on an online appeared first on My Assignment Tutor.

• • Assignment status: Resolved by our Writing Team
  • Source@PrimeWritersBay.com

GET THIS PAPER COMPLETED FOR YOU FROM THE WRITING EXPERTS  CLICK HERE TO ORDER 100% ORIGINAL PAPERS AT PrimeWritersBay.com

• • • • NO PLAGIARISM