Fair Fraud detection

End of Prototype phase

​

Group G1

Agenda

Dashboard Design

01

02

System Architecture

03

About Classifier

Dashboard Overview

A simulated, top-down, feedback based, continuous risk monitoring interface

• Simulated Flexibility

A playground that merchants could be able to simulate/compare their performance by adjusting provided parameters.

​

• Top-down Approach

From top-level metrics (fraud rate, revenue and chargeback costs) to detailed decision making explanation graph based on feature importance.

• Continuous Monitoring

Visualizations of the impact of historical decisions by metrics together with the evolutions of predicted risk scores of sensible features such as IP, Email and Account.

​

• Feedback Mechanism

Possibility to give feedback for wrongly predicted transactions. (Error feature will be available in Feature Store, ready to be served for next training iteration)

Add Screenshots of the Dashboard DEMO

BEFORE:

​

DB = 0.46

AFTER:

​

DB = 0.57

System Overview

First MVP

Incremental Random Forest with boosting

Jan-June Data

​

using updated weights on wrongly classified points from adyen predictions *

Feb-July Data

​

using updated weights on wrongly classified points from previous forest *

July

August

September

Mar-August Data

​

using updated weights on wrongly classified points from previous forest *

• Fraud probability of new incoming transaction would be based on soft voting weighted by forest f1 score.

​

• Feature contributions of new prediction would be calculated by weighted average by forest f1 score.

* Wrong points reported by the merchants/consumers as well as the feedback of the banks

About the Classifier

Questions