Actionable Objective Optimization for Suspicious Behavior Detection - - PowerPoint PPT Presentation

▶
actionable objective optimization
SMART_READER_LITE
LIVE PREVIEW

Actionable Objective Optimization for Suspicious Behavior Detection - - PowerPoint PPT Presentation

Feb 11, 2023 179 likes •584 views

Actionable Objective Optimization for Suspicious Behavior Detection on Large Bipartite Graphs Tong Zhao, Matthew Malir, Meng Jiang DM2 Laboratory Computer Science and Engineering University of Notre Dame Suspicious Behavior on Bipartite Graph

slide-1
SLIDE 1

Actionable Objective Optimization for Suspicious Behavior Detection

  • n Large Bipartite Graphs

Tong Zhao, Matthew Malir, Meng Jiang

DM2 Laboratory Computer Science and Engineering University of Notre Dame

slide-2
SLIDE 2

Suspicious Behavior on Bipartite Graph

  • Bot followers in social networks.
  • Bully buyers in e-commercial platforms.

Tong Zhao 2

Notice: Recently, there are some customers making improper evaluations and comments like posting ads or asking for cashbacks. Taobao.com has banned them from posting any comments.

slide-3
SLIDE 3

Suspicious Behavior on Bipartite Graph

  • Bot followers in social networks.
  • Bully buyers in e-commercial platforms.
  • Behavior: source users target users.
  • Source users: followers, buyers.
  • Target users: followees, sellers.

Tong Zhao 3

slide-4
SLIDE 4

Key Observation/Assumption

  • Fraudsters’ avoiding effort

forms dense blocks.

Tong Zhao 4

slide-5
SLIDE 5

Key Observation/Assumption

  • Fraudsters’ avoiding effort

forms dense blocks.

Tong Zhao 5

Follower seller

slide-6
SLIDE 6

Key Observation/Assumption

  • Fraudsters’ avoiding effort

forms dense blocks.

Tong Zhao 6

Follower seller

slide-7
SLIDE 7

Existing Methods

  • Find a dense subgraph.
  • Find the suspiciousness

vector 𝒗.

max

𝑣

𝐾(𝑩𝑡𝑣𝑐(𝒗)) 𝐾 𝐵𝑡𝑣𝑐 = 𝑓 𝑜𝑣 × 𝑜𝑤

Tong Zhao 7

(1) (2)

slide-8
SLIDE 8

Does it work?

  • Yes but NOT Actionable!

Tong Zhao 8

slide-9
SLIDE 9

Does it work?

  • Yes but NOT Actionable!
  • Serious consequence of false positive.

– An important email is thrown into spam. – A normal Twitter/Taobao account is banned.

  • Double check the reported suspicious users?
  • Heavy human labor.

Tong Zhao 9

Large size of data.

slide-10
SLIDE 10

What is actionable?

  • Blocklist function

Tong Zhao 10

slide-11
SLIDE 11

What is actionable?

  • Blocklist function
  • Blocking plug-ins.

Tong Zhao 11

Allow buyers whose average rating (AR) ≥ 0.92 to purchase items;

Blocking bullies – Settings

slide-12
SLIDE 12

What is actionable?

  • Blocklist function
  • Blocking plug-ins.

Tong Zhao 12

Seller “You cannot purchase if your AR is lower than 95%.” Screenshot of the buyer’s profile: “…. AR given by the buyer: 85.19% …” “Please use another account if you have.”

slide-13
SLIDE 13

Observation

Tong Zhao 13

Decision Massive Behavior Data

slide-14
SLIDE 14

Observation

Tong Zhao 14

Individual Experience

Decision Massive Behavior Data

slide-15
SLIDE 15

The Gap

Tong Zhao 15

Individual Experience

Decision Massive Behavior Data

Action Gap: Actionable Objectives

slide-16
SLIDE 16

Our Idea

Tong Zhao 16

Buyer’s average rating … … … …

slide-17
SLIDE 17

Our Idea

Tong Zhao 17

Buyer’s average rating … … … …

slide-18
SLIDE 18

Our Idea

Tong Zhao 18

Buyer’s average rating … … … …

slide-19
SLIDE 19

Our Idea

Tong Zhao 19

Buyer’s average rating … … … …

slide-20
SLIDE 20

Our Idea

  • Use platform’s big data.
  • Learn the best threshold for everyone.
  • Actionable Objective Optimization (AOO):

find the threshold vector 𝒘.

Tong Zhao 20

slide-21
SLIDE 21

AOO

Tong Zhao 21

Bij = ቊ1, 𝑗𝑔 𝐽𝑗𝑘 = 1 𝑏𝑜𝑒 𝑣𝑗 < 𝑤𝑘; 0, 𝑝𝑢ℎ𝑓𝑠𝑥𝑗𝑡𝑓.

(3)

slide-22
SLIDE 22

AOO

  • Calculate the indicator vectors.

Tong Zhao 22

𝑑(𝑣) = 𝐂 ∙ 𝟐𝑛 𝑡

𝑘 (𝑤) = ൝1,

𝑗𝑔 𝑑

𝑘 (𝑤) ≥ 𝛾(𝑤);

0, 𝑝𝑢ℎ𝑓𝑠𝑥𝑗𝑡𝑓. 𝑡𝑗

(𝑣) = ൝1,

𝑗𝑔 𝑑𝑗

(𝑣) ≥ 𝛾(𝑣);

0, 𝑝𝑢ℎ𝑓𝑠𝑥𝑗𝑡𝑓. 𝑑(𝑤) = 𝐂 ∙ 𝟐𝑜 (4) (7) (6) (5)

slide-23
SLIDE 23

AOO

  • Find the size and sum of the block.

Tong Zhao 23

𝑜𝑣 = 𝟐𝑜

𝑈 ∙ 𝑡(𝑣)

𝑜𝑤 = 𝟐𝑛

𝑈 ∙ 𝑡(𝑤)

𝑓 = 𝑡(𝑣)𝑈 ∙ 𝐂 ∙ 𝑡(𝑤) (8) (10) (9)

slide-24
SLIDE 24

AOO

  • Objective function that we want to maximize.

Tong Zhao 24

𝐾𝑒 𝒘 = 𝑓 𝑜𝑣 × 𝑜𝑤 = 𝑡(𝑣)𝑈 ∙ 𝐂 ∙ 𝑡(𝑤) (𝟐𝑜

𝑈 ∙ 𝑡(𝑣))(𝟐𝑛 𝑈 ∙ 𝑡(𝑤))

(12) (11)

slide-25
SLIDE 25

AOO

  • Find the partial derivatives with respect to 𝒘.

Tong Zhao 25

𝜖𝐾𝑒 𝜖𝑤𝑙 = 1 𝑜𝑣𝑜𝑤 𝜖𝑓 𝜖𝑤𝑙 − 𝑓 𝑜𝑣

2𝑜𝑤

𝜖𝑜𝑣 𝜖𝑤𝑙 − 𝑓 𝑜𝑣𝑜𝑤

2

𝜖𝑜𝑤 𝜖𝑤𝑙 𝜖𝑜𝑣 𝜖𝑤𝑙 = ෍

𝑗=1 𝑜

෍

𝑘=1 𝑛 𝜖𝑜𝑣

𝜖𝐶𝑗𝑘 𝜖𝐶𝑗𝑘 𝜖𝑤𝑙 = α2 ෍

𝑗=1 𝑜

𝑡𝑗

(𝑣) 1 − 𝑡𝑗 (𝑣)

𝐶𝑗𝑙(1 − 𝑕(𝑤𝑙 − 𝑣𝑗))

(13) (14) (15)

slide-26
SLIDE 26

AOO

Tong Zhao 26

𝜖𝑜𝑤 𝜖𝑤𝑙 = ෍

𝑗=1 𝑜

෍

𝑘=1 𝑛 𝜖𝑜𝑤

𝜖𝐶𝑗𝑘 𝜖𝐶𝑗𝑘 𝜖𝑤𝑙 𝜖𝑓 𝜖𝑤𝑙 = ෍

𝑗=1 𝑜

෍

𝑘=1 𝑛

𝜖𝑓 𝜖𝐶𝑗𝑘 𝜖𝐶𝑗𝑘 𝜖𝑤𝑙

(16) (17) (18)

= α2𝑡𝑙

(𝑤)(1 − 𝑡𝑙 (𝑤)) ෍ 𝑗=1 𝑜

𝐶𝑗𝑙(1 − 𝑕(𝑤𝑙 − 𝑣𝑗))

= α2 ෍

𝑗=1 𝑜

𝑡𝑗

(𝑣) 1 − 𝑡𝑗 (𝑣)

෍

𝑟=1 𝑛

𝐶𝑗𝑟𝑡𝑟

(𝑤) + 𝑜𝛽𝑡𝑙 (𝑤) 1 − 𝑡𝑙 (𝑤) ෍ 𝑞=1 𝑜

𝐶𝑞𝑙𝑡𝑞

(𝑣) + 𝑡𝑙 (𝑤) ෍ 𝑗=1 𝑜

𝑡𝑗

(𝑣)

(19)

slide-27
SLIDE 27

AOO

Tong Zhao 27

Each seller 𝐾𝑒 𝒘 and 𝜖𝐾𝑒 𝜖𝒘

slide-28
SLIDE 28

Experiments

  • On both synthetic datasets and real-world datasets.
  • Baselines:

– SpokEn (Prakash, et al., 2010) – CatchSync (Jiang, et al., 2014) – Fraudar (Hooi, et al., 2016) – Actionable version of each of them.

Tong Zhao 28

slide-29
SLIDE 29

Actionable Version Baselines

Tong Zhao 29

slide-30
SLIDE 30

Actionable Version Baselines

Tong Zhao 30

slide-31
SLIDE 31

Actionable Version Baselines

Tong Zhao 31

slide-32
SLIDE 32

Synthetic Data

Tong Zhao 32

slide-33
SLIDE 33

Experiment Results

Tong Zhao 33

Sparse blocks Dense blocks

slide-34
SLIDE 34

Experiment Results

  • Changing the

attack density.

  • Number of

blocks = 3.

Tong Zhao 34

slide-35
SLIDE 35

Real-word Dataset

  • Amazon reviews in 2015.
  • 4,552 users (buyers).
  • 6,347 products (sellers).
  • 231,600 ratings with reviews.

Tong Zhao 35

slide-36
SLIDE 36

Observations on Real-word Dataset

Tong Zhao 36

Buyers’ average rating distribution Word cloud of all reviews

slide-37
SLIDE 37

Observation on the Results

Tong Zhao 37

Word cloud of all reviews blocked by AOO Word cloud of all bad reviews not blocked by AOO

slide-38
SLIDE 38

Conclusions

  • Revisited the problem of suspicious behavior detection from

the perspective of individuals.

  • Proposed a novel Actionable Objective Optimization (AOO)

method that finds actionable solution of preventing fraud behaviors to happen.

  • Experimental results showed that AOO is effective and

efficient.

Tong Zhao 38

slide-39
SLIDE 39

Thank you!

  • Any questions?

Tong Zhao 39

slide-40
SLIDE 40

Synthetic Experiment Results

  • Quadratic time complexity.
  • 𝑃(𝑛𝑜𝑠𝑢).

– 𝑛: number of users. – 𝑜𝑠: number of ratings. – 𝑢: number of iterations.

Tong Zhao 40