An Auctioning Reputation System Based on Anomaly Detection

Shai Rubin, Mihai ChristodorescuVinod Ganapathy, Jonathon T. Giffin,

Louis Kruger, Hao Wang, Nicholas Kidd

Computer Sciences Dept.University of Wisconsin

Madison

2

Online Auctioning• Huge volume: eBay hosted 440,100,000 new

listings in Q2 2005• In this talk: trustworthiness of online auctioning • Why do we buy in an online auction?

A. to find a rare/collectable itemB. to find a bargain; commodity at a “good” price

• eBay financial report (expected 2005):– Clothing & Accessories --- $3.3 billion (2nd)– Consumer Electronics --- $3.2 billion (3rd)– Computers --- $2.9 billion (4th)

Data suggests that most people use eBay to find bargains

3

Finding a Bargain is Tricky

• Inherently untrustworthy environment: – Pseudonymous sellers– Pseudonymous buyers – Delivery? Warranty? Quality?

• Reputation system: a tool to establish trust

4

Finding a Bargain is Tricky

• eBay’s reputation system provides little help– Based on feedback: vulnerable to “poisoning” attack

5

Finding a Bargain is Tricky

• eBay’s reputation system provides little help– Based on feedback: vulnerable to “poisoning” attack – Does not provide information on price

6

Finding a Bargain is Tricky

• eBay’s reputation system provides little help– Based on feedback: vulnerable to “poisoning” attack – Does not provide information on price– Does not differentiate among the majority of sellers

60%

70%

80%

90%

100%

0% 20% 40% 60% 80% 100%

% of sellers (1545 sellers with more than 10 auctions)

90% of sellers:Positive feedback > 97.3%

50% of sellers:Positive feedback > 99.4%

% of positivefeedback

7

Goals• Alice—a buyer, Bob—a seller• Develop a trustworthy mechanism that helps Alice:

– Achieve her goal: what are the chances that Alice can find a bargain in Bob’s auctions?

– Warn her from fraudulent activities: are the prices in Bob’s auctions artificially inflated?

– Provide her assurance against poisoning attack: why should Alice trust the mechanism?

8

Contributions

• A reputation system that helps buyers avoid sellers who seem to be inflating prices– Formulated the “seem to be inflating prices” as an

anomaly detection problem – Business level anomaly detection: the basic events are

auctions, bidding. – Behavioral system: based on how human behave/act

rather than on people feedback. • Only a first step, some goals still ahead

9

Outline

• Motivation: find a bargain and avoid fraud• Contributions: anomaly detection system to

identify price inflating sellers:– The N model– The M model– The P model

• Case studies

10

Auctioning 101• Pseudonymous sellers and bidders • Auctions end after a predefined time (e.g., 7 days)• Highest bid wins • Seller sets minimum starting bid• Shilling: a group of bidders that place fake bids to

inflate the final price

11

Methodology• Collect data from eBay

– three weeks of data in the category: Laptop Parts & Accessories

– 127,815 auctions, 12,331 sellers, – 604 high-volume sellers: posted more than 14 auctions

controls 60% of the market• Use statistical model to predict seller behavior

– 95% of the sellers are “normal”– 5% are abnormal, or suspicious

12

Step 1: Average Number of Bids• What is an indication that prices are high?

– high number of bids • Goal: identify sellers with abnormally high

number of bids

0 2 4 6 8 10 12 14 160

5

10

15

20

25

30

35

% o

f hig

h−vo

lum

e se

llers

Average number of bids per auction

5%95%

5%95%

Average bids per auction

% o

f hig

h-vo

lum

e se

llers

• 95% of high-volume sellers have less than 7 bids per auction

•Model is insensitive to supply: number of auctions posted by a seller

13

Step 1: The N Model

0 2 4 6 8 10 12 14 16

200

400

600

800

1000

1200

1400

1600

Average number of bids per auction

Num

ber

of a

uctio

ns fo

r a

selle

r

10260

N normalN abnormalNormal line

Num

ber o

f auc

tions

for s

elle

r

Suspicious:5% of high-volume sellers

Correlation: many auctions implies low number of bids

Suspicious seller: one that posts many auctions and still attracts many bids

Average bids per auction

14

Outline

• Motivation: find a bargain and avoid fraud• Contributions: anomaly detection system to

identify price inflating sellers:– The N model: a seller is suspicious if they post many

auctions that attract many bids– The M model– The P model

• Reputation example

15

Step 2: Average Minimum Starting Bid• Legitimate explanation for high number of bids:

low minimum starting bid • Goal: identify sellers with abnormally high

number of bids and high minimum bid• Problem: how do you know that the minimum bid

is high?

dwinning_bidminimum_bidwinning_bi −Relative minimum

bid (RMB) =

16

Step 2: The M Model

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

2

4

6

8

10

12

14

16

10260

Average RMB for seller

Ave

rage

# o

f bid

s in

sel

ler’s

auc

tions

M normalM abnormalPolynomial fitNormal line

Ave

rage

# o

f bid

s in

sel

ler’s

auc

tions

Average relative minimum bid for seller

Suspicious: 5% of high-volume sellers

Correlation: low minimum starting bid implies many bids

M suspicious seller: starts with high minimum bid and attracts many bids

M+N suspicious seller: posts many auctions, attracts many bids, starts with high minimum bid

17

Step 3: Bidders’ Profile of a Seller • Fraudulent explanation for high number of bids:

shilling • Goal: identify group of bidders that repeatedly

bid and lose in a seller’s auctions• Suspicious seller:

– N: sellers with abnormally high number of bids and– M: high starting bid and– P: has a group of bidders that repeatedly bid and lose

18

Bidder Presence Curve

0 10 20 30 40 50 60 70 80 90 1000

20

40

60

80

100P

artic

ipat

ed a

uctio

ns (

%)

Bidders (%)

Bidder presence

0 10 20 30 40 50 60 70 80 90 1000

20

40

60

80

100

Bidders (%)

Par

ticip

ate

or w

in (

%)

Bidder presenceBidder wins

Bidders %

5% of bidders in this seller’s auctions participated in 80% of the auctions

Participated auctions %

19

Bidder Presence/Win Curves

0 10 20 30 40 50 60 70 80 90 1000

20

40

60

80

100P

artic

ipat

ed a

uctio

ns (

%)

Bidders (%)

Bidder presence

0 10 20 30 40 50 60 70 80 90 1000

20

40

60

80

100

Bidders (%)

Par

ticip

ate

or w

in (

%)

Bidder presenceBidder wins

Bidders %

Participated auctions %

the same 5% won only 10% of the auctions

Bidders %

Participatedor won

auctions %

5% of bidders in this seller’s auctions participated in 80% of the auctions

20

0 10 20 30 40 50 60 70 80 90 1000

20

40

60

80

100

Par

ticip

ated

auc

tions

(%

)

Bidders (%)

C Normal lineBidder presence

0 10 20 30 40 50 60 70 80 90 1000

20

40

60

80

100

Bidders (%)

Par

ticip

ate

or w

in (

%)

(d) Cumulative bidder presence and bidder wins curves

C Bidder presenceBidder wins

Bidder Presence/Win Curves(Normal case)

Bidders %

Participatedor won

auctions %

10% of the bidders participated in 20% of the auctions and won 20% of the times

21

Outline

• Motivation: find a bargain and avoid fraud• Contributions: anomaly detection system to

identify price inflating sellers:– The N model: a seller is suspicious if they post many

auctions that attract many bids – The M model: a seller is suspicious if they attract

many bids and start with high minimum bid– The P model: a seller is suspicious if they have a

group of bidders that repeatedly participate and lose• Reputation example

22

Reputation Example: Seller 10260

0 2 4 6 8 10 12 14 16

200

400

600

800

1000

1200

1400

1600

Average number of bids per auction

Num

ber o

f auctions for a

seller

10260

N normalN abnormalNormal line

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

2

4

6

8

10

12

14

16

10260

Average RMB for seller

Average #

of bid

s in s

eller’s

auctions

M normalM abnormalPolynomial fitNormal line

0 20 40 60 80 1000

10

20

30

40

50

60

70

80

90

100

Bidders (%)

Pa

rtic

ipa

te

o

r w

in (%

)

Bidder presenceBidder wins

N M P

Bidders %Average relative minimum bid for seller

Average bids per auction

23

Results Summary

• 54 sellers classified as abnormal with respect to at least one model

• 3 sellers classified as abnormal with respect to all three models

• No confirmed fraud

N M

P

3

2

21

24 13

9

24

Summary

• Trust: do we get what we expected?• Reputation system as anomaly detection

– Attempt to identify price inflation– Work at the business level– Consider poisoning attack (see paper)

Thank you. Questions?