hendra suryanto - achieving enterprise level value from machine based learning - futuredata 2017
TRANSCRIPT
Achieving enterprise level value from machine based learningHendra SuryantoChief Data ScientistRich Data Corporation
Presented in
Data | Innovation | Disruption
• Mobile Lending Platform utilizing our Credit Decision Engine that leverages social media and other data sources
• Management and board with extensive global financial services experience.
• Enrich traditional consumer data with Social Media data
• Turn data into value through innovative business models
• Focusing on consumer profiling, behaviour data and predictive analytics
• Taking technology incubated in Australia to Asia
• RDC platforms deployed in Australia, Singapore, Canada and Vietnam: 6 successful implementations in last 18 months.
• Signed agreements with large enterprises in China, Vietnam, and US.
Market MomentumAI & Social Media Enabled Fintech
RICH
DATA CORP
ORA
TION
2
Machine Learning – Case Studies
• Marketing: next best offers • Fraud: credit card fraud • Customer journey analytics: churns • Risk: credit scoring
Various feedback lags – various learning speed
Machine Learning – Case Studies
• Marketing: next best offers (1 minute – 1 month)• Fraud: credit card fraud (1 hour – 24 hours)• Customer journey analytics: churns (1 month – 6 month)• Risk: credit scoring (2 months – 12 months)
Various feedback lags – various learning speed
Case Study: credit risk on unsecured lending
A lender in Canada would like to automate their credit decisioning
Machine Learning
Human decision Machine decision
Case Study: credit risk on unsecured lending
Prototype Productionise Operationalise
Case Study: credit risk on unsecured lending
Prototype Productionise Operationalise
• Data scientist• One‐off /tactical
• data extract• feature
engineering• modelling• scoring &
decisioning
• Enterprise Architect• Automation
• data integration
• data & feature quality monitoring
• Business Team (Risk)• Process & people
• training• governance• scoring &
decisioning update
4 weeks 8 weeks 12 weeks
From raw data to business value ($$$)
• What is it?• Why we need it?• How do we delivering it?
Even
t tab
le: Entity
Attrib
ute Va
lue
Timestamp (EAV
T)
Flat Table (Entities x Features Cross Table)
Descriptive Analytics
Predictive Analytics
Prescriptive Analytics
Raw/Binary DataFeature Engineering
Analytics
Business Outcomes (in Dollars)
From raw data to business value ($$$)
• What is it?• Why we need it?• How do we delivering it?
Learning from event based data (raw data)
data collection time
Loan defaulter
Good customer
Entity: Customer
Attribute‐ValueCall Reason=Complaint
Timestamp
Timestamp is important to ensure we use historical data to predict future outcomes
Feature engineering – the missing link
• Feature selection• Feature extraction (e.g. NLP,
image processing)• Feature construction (e.g.
business logic, formulae)
Feature engineering – can Deep Learning help?
Domain Expert
Building Scoring Model
Building Decision Engine
Continuously learning from users and new data. Any feedback will be used for future learning.
Hierarchy of Features
Machine Learning
Domain Expert
Users
UsersA
B
New data
Continuous Learning
14
19,117 Applications were processed. $10 million loan issued. 10% increase in average loan size. Bad debt rate is 5% for the under banked segment.
Gini 0.53
Gini results from all training and test data.
Gini 0.52
Gini results after excluding Ageand Work Type to comply with Canadian Regulation
Gini 0.40
Gini in production is lower than testing due to data quality issues.
After retraining, the Gini is raised to our expected threshold. It improved also due to new features.
Gini 0.55
Learning in Real LifeCase Study: Canada
Machine Learning – Achieving Enterprise Value
• Expand market share (customer acquisition)
• Increase profit per customer• Reduce cost
Machine Learning
• Prototype• Productionise ‐ automation• Operationalise ‐ continuous
learning
Singapore
One Raffles Place #34‐04 Tower 1Singapore 048616
Australia
802, 8 West Street North SydneyNSW 2060
F facebook.com/richdataco
T @richdataco
W www.richdataco.com
China
Room 509, 5th Floor, Building 2Xunmei Technology PlazaYeuhai Street, Nanshan DistrictShenzhen, China