Sponsored Search

thumbpinchInternet και Εφαρμογές Web

18 Νοε 2013 (πριν από 4 χρόνια και 1 μήνα)

74 εμφανίσεις

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


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

Bidder

Bid for

q
1

Bid for
q
2

Budget

b
1

C
1

+


C
1

C
1

b
2

C
1

0

C
1

b
3

C
1

-



C
1

-



2
C
1

Allocation

Shown
for
q
1

Shown
for
q
2

Total
Revenue

Greedy

b
1

b
3

C
1

+


Optimal

b
2

b
1

2C
1

-



Reserve price is


Problem Definition


Queries
Q = {q
1
, q
2
, q
3
, ..., q
N
}


Bidders
B = {b
1
, b
2
, b
3
, ..., b
M
}


Bidding state
A(t);

A
ij
(t)
is
j
’s bid


for
i
-
th
query


d
j

is
j
’s daily budget



v
i

is estimate of query frequency


L
i

= {j
p

: j
p



B, p = 1, ..., P
i
}


L
i
k

= {j
i
k

: j
i
k



L
i
, l ≤
L
i
k

≤ P}

Ranking and revenue


Bid
-
ranking
-



Revenue
-
ranking
-



So, for slate
k
,


Price per click:


Independent click through rates


Revenue
-
per
-
search:




Total revenue:

Bidder’s cost


Total spend for
j
:

Linear program


Queries
i =
1, ...,
N


Bidders
j =
1, ...,
M


Slates
k =
1, ...,
K
i



Data:
d
j
, v
i
, c
ijk
, r
ik


Variables:
x
ik


Constraints:



Budget:



Inventory:

Objective function


Maximize revenue:



Value objective:



Clicks objective:


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
i
th

keyword


Profit if


Maximize

How to maximize?


If small number of bidders for a query,
enumerate all legal subsets
L
i
k
, find
maxima, see if adding increases profit


Otherwise, use algorithm described in
another paper

tigerdirect.com

?

nextag.com

priceline.com

ebay.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
f
ik

= x
ik
/v
i


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

Gains when revenue is maximized

Impact on bidders

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