1 of 10

Applied M/L for real-time FairPlay against Fraud

​

Aditya Prasad Narisetty

@adityaprasadn

​

2 of 10

Challenges @Dream11,

India’s largest fantasy sports platform

​

Referral Bonus Abuse

Promotions Abuse

User colluding

3 of 10

Layers of Shield

​

  1. User Registration
    1. User Fingerprinting
    2. Referral channel
    3. All users in the same referral family
    4. Cross family user similarity?!
  2. Winnings Withdrawal
    • Final layer of verification
    • Contests played of the user
    • Violated any FairPlay conditions

4 of 10

Training Data

​

  1. User Fingerprinting
    1. User Profile
      1. Incremental bonus for every step of KYC
      2. Email, Name, ZipCode, DateOfBirth, PasswordHash, Referral Family, TimeOfCreation, ID
    2. System Profile
      • Cookies, UserAgent, IP address, isRooted, MacId
  2. Game Play
    • Contests played by the user
    • Something fishy in the contests?
  3. 6 months of data
    • Manual tagging
    • Human bias

5 of 10

Is this Registration an FPV?

​

​

Classification

Classify a new registration as FPV

Linear model?

Need a relative score for prioritisation

Features

Evaluate similarity of User and System profile?

Combine similarity scores for an aggregate metric to use as a classifier

​

6 of 10

Is this Withdrawal an FPV?

​

​

Classification

Classify a new user withdrawal as FPV

Need a relative score for prioritisation

Features

Incorporate referral FPV score

Similarity of users’ team w.r.t other teams in a contest

Patterns similar to abusive users

User colluding : similar or orthogonal teams in the same contest

Contest with users of the same referral family

7 of 10

Modelling

Logistic Regression

Pre-trained weights for every feature

Add logits from compound features(sigmoid normalized)

Final Score used a classifier

​

​

UserID

NameSim

EmailSim

ZipSim

IPSim

UserAgentSim

​

UserSim

SystemSim

​

FinalFPV

532XX4

0.3

0.8

0.8

0.2

0.1

0.9

0.05

0.86

532XX5

0.1

0.03

0.3

0.8

0.9

0.2

0.95

0.92

8 of 10

Pipeline

Engineering:

Clickstream data

Kafka for scaling upto 100k+ events/sec

Spark Aggregator

Model Deployment:

Stacked models of Linear and XGBOOST decisions

Scikit Pipeline

Django API

ElasticSearch monitoring & alerting

​

9 of 10

Business impact & What lies ahead?

​

Impact:

Registration FPV 98.3% ROC-AUC

Withdrawal FPV 94.2% ROC-AUC

Decreased human workload by 90+%

Further:

Likelihood of cross family colluding

Proactive user activity monitoring

Remove human bias

Adaptive learning

10 of 10

Thank You

​

@AdityaPrasadN

@D11Engg