Sponsored Search - California Institute of Technology

Report
Sponsored Search
Cory Pender
Sherwin Doroudi
Optimal Delivery of Sponsored Search
Advertisements Subject to Budget Constraints
Zoe Abrams
Ofer Mendelevitch
John A. Tomlin
Introduction
Search engines (Google, Yahoo!, MSN)
auction off advertisement slots on
search page related to user’s keywords
 Pay per click
 Earn millions a day through these
auctions

– Auction type is important
Sponsored search parameters
Bids
 Query frequencies

QuickTime™ and a
TIFF (LZW) decompressor
are needed to see this picture.
– Not controlled by advertisers or search engine
– Few queries w/ large volume, many with low
volume
Advertiser budgets
 Pricing and ranking algorithm

Solution
 Focus on small subset of queries
– Predictable volumes in near future
– Constitute large amount of total volume
Sponsored search parameters

Bids
 Query frequencies
 Advertiser budgets
– Controlled by advertisers

Pricing and ranking algorithm
– Generalized second price (GSP) auction
– Rankings according to (bid) x (quality score)
– Charged minimum price needed to maintain rank

Goal: take these parameters into account,
maximize revenue
Motivating example
Reserve price is 
Bidder
Bid for q1 Bid for q2 Budget
b1
C1 + 
C1
C1
b2
C1
0
C1
b3
C1 - 
C1 - 
2 C1
Allocation Shown
for q1
Greedy
b1
Shown
for q2
b3
Total
Revenue
C1 + 
Optimal
b1
2C1 - 
b2
Problem Definition
Queries Q = {q1, q2, q3, ..., qN}
 Bidders B = {b1, b2, b3, ..., bM}
 Bidding state A(t); Aij(t) is j’s bid
for i-th query
 dj is j’s daily budget
 vi is estimate of query frequency
 Li = {jp : jp  B, p = 1, ..., Pi}
 Lik = {jik : jik  Li, l ≤ Lik ≤ P}

Ranking and revenue






QuickTime™ and a
TIFF (LZW) decompressor
are needed to see this
picture. and a
QuickTime™
Bid-ranking TIFF (LZW) decompressor
Revenue-ranking areQuickTi
neededme™
to see
andthis
a picture.
TIFF (LZW) decompressor
So, for slate k,
are needed to see this pi cture.
QuickTime™ and a
TIFF (LZW) decompressor
Price per click: are needed to see this picture.
Independent click through rates
Qu ickTime ™ a nd a
TIFF (L ZW) dec omp ress or
are nee ded to s ee th is p ictur e.
Revenue-per-search:
Total revenue:
Qui ckTime™ and a
TIFF ( LZW) decompressor
are needed to see this pi cture.
QuickTime™ and a
TIFF (LZW) decompressor
are needed to see this picture.
Bidder’s cost

Total spend for j:
Quic kT ime™ and a
T IFF (LZW) dec ompres sor
are needed to s ee this pi cture.
QuickTime™ and a
TIFF (LZW) decompressor
are needed to see this picture.
Linear program
Queries i = 1, ..., N
 Bidders j = 1, ..., M
 Slates k = 1, ..., Ki
 Data: dj, vi, cijk, rik
 Variables: xik
 Constraints:

– Budget:
– Inventory:
QuickTime™ and a
TIFF (LZW) decompressor
are needed to see this picture.
QuickTi me™ and a
T IFF (LZW) decompressor
are needed to see thi s pi cture.
Objective function



Maximize revenue:
Value objective:
Clicks objective:
Quic kT ime™ and a
T IFF (LZW) dec ompres sor
are needed to s ee this pi cture.
QuickTime™ and a
TIFF (LZW) decompressor
are needed to see this picture.
QuickTime™ and a
TIFF (LZW) decompressor
are needed to see this picture.
Column Generation
Each column represents a slate
 Could make all possible columns

– But for each query, exponential in number
of bidders
Start with some initial set of columns
 j: Marginal value for j’s budget
 i: Marginal value for ith keyword
QuickTime™ and a
(LZW) decompressor
 Profit if areTIFF
needed to see this picture.
 Maximize

QuickTi me™ and a
TIFF (LZW) decom pressor
are needed to see this pi cture.
How to maximize?
If small number of bidders for a query,
enumerate all legal subsets Lik, find
maxima, see if adding increases profit
 Otherwise, use algorithm described in
another paper

ebay.com
tigerdirect.com
QuickTime™ and a
TIFF (LZW) decompressor
are needed to see this picture.
QuickTime™ and a
TIFF (LZW) decompressor
are needed to see this picture.
?
nextag.com
priceline.com
Summary (so far)






Various bidders vying for spots on the slate
for each query
Constrained by budget, query frequencies,
ranking method
Solve LP for some initial set of slates
Check if profit can be made by adding new
slates
Re-solve LP, if necessary
Can be applied to maximize revenue or
efficiency
Simulation Methodology

Compare this method to greedy algorithm
– For greedy, assign what gets most revenue at the
time; when bidder’s budget is reached, take them
out of the pool

Used 5000 queries
 For 11 days, retrieved hourly data on bidders,
bids, budgets
 To determine which ads appear, assign based
on frequencies fik = xik/vi
 After each hour, see if anyone has exceeded
budget
Simulation Results
Current method better than greedy
method, when optimizing over revenue
or efficiency
 Larger gain for revenue when revenue
optimized
 Revenue and efficiency are closely tied

Gains when efficiency is maximized
QuickTime™ and a
TIFF (LZW) decompressor
are needed to see this picture.
Gains when revenue is maximized
QuickTime™ and a
TIFF (LZW) decompressor
are needed to see this picture.
Impact on bidders
QuickTime™ and a
TIFF (LZW) decompressor
are needed to see this picture.
Limitations
Illegitimate price hikes for other bidders
if one person exceeds budget in middle
of hour
 Assumption that expected number of
clicks are correct
 For the purposes of the simulation,
expect these to affect greedy and LP
optimization similarly

Future work

Focus on less frequent queries
– Frequencies harder to predict
– Some work has been done (doesn’t incorporate
pricing and ranking)

Keywords with completely unknown
frequencies
 Parallel processing for submarkets
 Investigate how advertisers might respond to
this method
– Potential changes in reported bids/budgets

similar documents