Intended for Knowledge Sharing only
Predictive Analytics as a Product
Feb 2017
Intended for Knowledge Sharing only
Disclaimer:
Participation in this summit is purely on personal basis and is not meant to represent VISA’s position on
this or any other subject and in any form or matter. The talk is based on learning from work across
industries and firms. Care has been taken to ensure no proprietary or work related information of any
firm is used in any material.
Intended for Knowledge Sharing only
Quick recap of what it is
Intended for Knowledge Sharing only
Data Scientist, eh…
3
Intended for Knowledge Sharing only
Quick recap of what it is
Intended for Knowledge Sharing only
FEELS LIKE A ROCKSTAR, DOESN’T IT?
4
http://modernservantleader.com/servant-leadership/narcissism-kills-morale/
Intended for Knowledge Sharing only
Quick recap of what it is
Intended for Knowledge Sharing only
..BUT A KANYE & NOT COLDPLAY
5
https://imgflip.com/memegenerator/7064654/Kanye-West
Intended for Knowledge Sharing only
Quick recap of what it is
Intended for Knowledge Sharing only
So what happened?
6
Intended for Knowledge Sharing only
Intended for Knowledge Sharing only
SOME CHALLENGES
7
Unrealistic expectations on RoI.
Operates in siloes, not complemented by user research/other internal or
external data/experimentation results.
Field testing & iterative development still predominantly offline.
Deployment, Post Deployment management & monitoring expensive. Not
easy to turn on/off, tweak, flip, scale.
Predictions driven significantly by historical trends and relationships.
Expectations modeled as simulations.
Intended for Knowledge Sharing only
Quick recap of what it is
Intended for Knowledge Sharing only
Explain it a bit more...
8
COMPLICATION 1: PREDICTIVE ANALYTICS IS INTRICATE & COMPLEX
Intended for Knowledge Sharing only 9
Objective
Translation to Analytical
Framework
Data Collection and
Preparation
Analysis, Validation &
Verification
Actionable insights and
impact sizing
A/B Testing
Rollouts
• Understand need, fit with Strategic needs, actionability, stakeholders buy-in, engineering
RoI, project management
• Decide on the Analytical methodology based on nature of the problem, dependent
variable, frequency, sample, time, required precision, actionability
• Hypothesized driver list
• Data Collection: Internal & external sourcing
• Data Preparation: Blending, aggregations
• Data Transformations: Outlier, Missing, math transformation, interactions, redundancy
treatments, variable selections
• Sampling methodology & split
• Model development and validation: In-time, Out-of-time
• Stand alone, ensemble
• Performance diagnostics & cross check with other sources
• Recommendations, impact sizing, cross leverage scores
• Field Testing (Champion vs. Challenger)
• Iteration plan based on user feedback (VOC), performance
• Model deployment, post deployment monitoring & management
• Integration with Product Line– New product,
1
2
3
4
5
6
7
Description
COMPLICATION 2: MULTIPLE AUDIENCE, PRIORITIES, DEPENDENCIES
Intended for Knowledge Sharing only 10
Objective
Translation to Analytical
Framework
Data Collection and
Preparation
Analysis, Validation &
Verification
Actionable insights and
impact sizing
A/B Testing
Rollouts
• Analyst & Stakeholder
• Analyst, Data Instrumentation, Data
Manager, Stakeholder
• Analyst, Data Instrumentation, Data
Manager
• Analyst
• Analyst, Stakeholder, Cross Functional
team, Leadership
• Analyst, Experimentation Team, User
Researcher, Developer, Stakeholder
• Analyst, Developer, Stakeholder,
Leadership
1
2
3
4
5
6
7
Who does it?
• Agile and may undergo iteration
• Changes in Strategic goals, newer
initiatives, releases, discoveries, reorgs
• Sourcing/Blending challenges: Data
handovers between systems, blending
challenges
• Scalability/automation
• Data movements/latencies/
teams/approvals
• Evolution of hypotheses, data
changes/errors, success criteria
• Competing priorities, data movements,
Scenario Simulations
• Success criteria, integration with
research/testing tools, iterations
• Integration with host systems,
engineering investment, model
tweaking, monitoring, customization
Key Challenges
COMPLICATION 3: OUTPUT OF ONE CAN BE INPUT/ADDITION TO ANOTHER
Intended for Knowledge Sharing only 11
Behavioral
Merchant
Performance
Clickstream/
Ops
Campaign
Performance
VOC/Social/
CRM
• Probability of Engagement/LTV
Growth/Churn/Loyalty
• Life event changes
• Product/Price Migrations
• Probability of Growth/Churn
• Next Best Product/Offer
• Network partners
• Conversion Rate Optimization
• Server Response Times
• Time to Purchase
• Campaign Responses
• Next Best Product/Offer
• Cross Channel target
• Promoter/Detractor & drivers
• Brand Appeal
• Theme/entity of engagement
Data Lake:
Enriched with
predictions
e.g., Uber’s cross
sell platform,
Google Calendar,
VDP
COMPLICATION 4: REAL DECISION MAKING NEEDS ADDITIONAL REASONING BEYOND
ANALYTICS
Intended for Knowledge Sharing only 12
Analytics provides insights into “actions”, Research context on “motivations” & Testing
helps verify the “tactics” in the field and everything has to be productized…
Strategy
Data
Tagging
Data
Platform
Reporting
Analytics
Research
Data
Products
Iterative
Loop Why such complexity?
Focus on Big Wins
Reduced Wastage
Quick Fixes
Adaptability
Assured execution
Learning for future
initiatives
Optimization
Intended for Knowledge Sharing only
Intended for Knowledge Sharing only
COMPLICATION 5: DEMANDS ON PREDICTIVE ANALYTICS HAVE INCREASED
13
Predictive
Analytics
Behavioral
Analytics
What are the
customers doing?
Voice of
Customer
What are the
customers
telling you?
Platform
Performance
How are you
delivering? Competitive
Are the
customers
buying
elsewhere?
Social Listening
How are
customers
discussing you?
…aaanddd Better, Faster, Cheaper, Monetizable
Intended for Knowledge Sharing only
Quick recap of what it is
Intended for Knowledge Sharing only
So, what do we need then?
14
Intended for Knowledge Sharing only
• Extensible
• Scalable
• Flexible
• Easy to integrate with
other techniques
Intended for Knowledge Sharing only
HIGH LEVEL SUMMARY OF NEEDS: MODULAR, SHAREABLE & MONETIZABLE
15
keywordsuggest.org Iconfinder WebPT
• Documentation
• Governance
• Integration with
project management
tools (collaboration)
• Security/Privacy
Management
• Value Abstraction
• API-able
Modular Shareable Monetizable
Intended for Knowledge Sharing only
Quick recap of what it is
Intended for Knowledge Sharing only
Potential Solutions
16
TWO DEPLOYMENT SOLUTIONS- PMML & PFA
Intended for Knowledge Sharing only 17
Data Mining Group an independent Vendor Led Consortium that develops Data Mining
Standards has come up with PMML (Predictive Model Mark Up Language) and PFA (Portable
Format for Analytics)
http://www.kdnuggets.com/2016/01/portable-format-analytics-models-production.html
http://dmg.org/
https://www.ibm.com/developerworks/library/ba-predictive-analytics4/ba-predictive-analytics4-pdf.pdf
https://www.ibm.com/developerworks/library/ba-ind-PMML1/
http://www.kdnuggets.com/faq/pmml.html
https://journal.r-project.org/archive/2009-1/RJournal_2009-1_Guazzelli+et+al.pdf
PMML PFA
File XML JSON & YAML
Maturity Mature but expanding Evolving
Nesting/Customization Model Parameters
Control Structures (Type System
of Model Parameters & data -
Callback function allowed)
Flexibility
Standard across most
scoring engines (better
than custom code)
More flexible than PMML but
safer than Custom Code
Scope
Data prep, Modeling,
Scoring, Sharing
+Pre/Post processing,
enforced memory model
PMML PROJECTS
Intended for Knowledge Sharing only 18
http://data-informed.com/pmml-puts-big-data-to-work/
POSITIONING OF PFA
Intended for Knowledge Sharing only 19
http://data-informed.com/pmml-puts-big-data-to-work/
Intended for Knowledge Sharing only
Quick recap of what it is
Intended for Knowledge Sharing only
Why this, Why now, why here?
20
Intended for Knowledge Sharing only
Intended for Knowledge Sharing only
BIGGER TRENDS THAT ARE SHAKING UP THE ANALYTICS WORLD FROM INSIDE OUT…
21
Demand Pressures: Complexity and nature of problems and their solutions,
type of audience & consumption framework evolving
Monetization opportunities- Direct, Indirect, Recurring
Artificial Intelligence, IoE and “Smart”ening of devices/systems faster than
expected.
Evolution of input data sources and integration of multiple insights sources
into decision making (A/B Testing, Research, Predictions/Scores from other
models)
Evolution from Service to Product to Platform (Build Once, Use
Everywhere)
…APIs are eating up our world
Intended for Knowledge Sharing only
Quick recap of what it is
Intended for Knowledge Sharing only
The parting words…
22
SUMMARY
Intended for Knowledge Sharing only
Predictive Analytics has stopped being “one-off competitive edge project
exercise” – it’s a necessary survival initiative for organizations
Scale, complexity, breadth of needs (including Monetization) demand
Platform approach.
“Build Once, Use Everywhere” -consumption of predictive analytics outputs
need to be easy to use, integrate, re-use/collaborate across multiple
initiatives
As everything becomes Productized via APIs, together they can become a
business problem solving ANI
23
Streaming Analytics is quickly evolving into Streaming Predictive Analytics
Intended for Knowledge Sharing only
Quick recap of what it is
Intended for Knowledge Sharing only
Appendix
THANK YOU!
Intended for Knowledge Sharing only
Would love to hear from you on any of the following forums…
https://twitter.com/decisions_2_0
http://www.slideshare.net/RamkumarRavichandran
https://www.youtube.com/channel/UCODSVC0WQws607clv0k8mQA/videos
http://www.odbms.org/2015/01/ramkumar-ravichandran-visa/
https://www.linkedin.com/pub/ramkumar-ravichandran/10/545/67a
RAMKUMAR RAVICHANDRAN
Intended for Knowledge Sharing only
Disclaimer:
Participation is purely on a personal basis and does not represent VISA,Inc. in any form or matter. The
talk is based on learning from work across industries and firms. Care has been taken to ensure no
proprietary or work related info of any firm is used in any material.
Director, Insights at Visa, Inc.
Enable Decision Making at the
Executives/ Product/Marketing level via
actionable insights derived from Data.
RAMKUMAR RAVICHANDRAN

Predictive Analytics as a Product

  • 1.
    Intended for KnowledgeSharing only Predictive Analytics as a Product Feb 2017
  • 2.
    Intended for KnowledgeSharing only Disclaimer: Participation in this summit is purely on personal basis and is not meant to represent VISA’s position on this or any other subject and in any form or matter. The talk is based on learning from work across industries and firms. Care has been taken to ensure no proprietary or work related information of any firm is used in any material.
  • 3.
    Intended for KnowledgeSharing only Quick recap of what it is Intended for Knowledge Sharing only Data Scientist, eh… 3
  • 4.
    Intended for KnowledgeSharing only Quick recap of what it is Intended for Knowledge Sharing only FEELS LIKE A ROCKSTAR, DOESN’T IT? 4 http://modernservantleader.com/servant-leadership/narcissism-kills-morale/
  • 5.
    Intended for KnowledgeSharing only Quick recap of what it is Intended for Knowledge Sharing only ..BUT A KANYE & NOT COLDPLAY 5 https://imgflip.com/memegenerator/7064654/Kanye-West
  • 6.
    Intended for KnowledgeSharing only Quick recap of what it is Intended for Knowledge Sharing only So what happened? 6
  • 7.
    Intended for KnowledgeSharing only Intended for Knowledge Sharing only SOME CHALLENGES 7 Unrealistic expectations on RoI. Operates in siloes, not complemented by user research/other internal or external data/experimentation results. Field testing & iterative development still predominantly offline. Deployment, Post Deployment management & monitoring expensive. Not easy to turn on/off, tweak, flip, scale. Predictions driven significantly by historical trends and relationships. Expectations modeled as simulations.
  • 8.
    Intended for KnowledgeSharing only Quick recap of what it is Intended for Knowledge Sharing only Explain it a bit more... 8
  • 9.
    COMPLICATION 1: PREDICTIVEANALYTICS IS INTRICATE & COMPLEX Intended for Knowledge Sharing only 9 Objective Translation to Analytical Framework Data Collection and Preparation Analysis, Validation & Verification Actionable insights and impact sizing A/B Testing Rollouts • Understand need, fit with Strategic needs, actionability, stakeholders buy-in, engineering RoI, project management • Decide on the Analytical methodology based on nature of the problem, dependent variable, frequency, sample, time, required precision, actionability • Hypothesized driver list • Data Collection: Internal & external sourcing • Data Preparation: Blending, aggregations • Data Transformations: Outlier, Missing, math transformation, interactions, redundancy treatments, variable selections • Sampling methodology & split • Model development and validation: In-time, Out-of-time • Stand alone, ensemble • Performance diagnostics & cross check with other sources • Recommendations, impact sizing, cross leverage scores • Field Testing (Champion vs. Challenger) • Iteration plan based on user feedback (VOC), performance • Model deployment, post deployment monitoring & management • Integration with Product Line– New product, 1 2 3 4 5 6 7 Description
  • 10.
    COMPLICATION 2: MULTIPLEAUDIENCE, PRIORITIES, DEPENDENCIES Intended for Knowledge Sharing only 10 Objective Translation to Analytical Framework Data Collection and Preparation Analysis, Validation & Verification Actionable insights and impact sizing A/B Testing Rollouts • Analyst & Stakeholder • Analyst, Data Instrumentation, Data Manager, Stakeholder • Analyst, Data Instrumentation, Data Manager • Analyst • Analyst, Stakeholder, Cross Functional team, Leadership • Analyst, Experimentation Team, User Researcher, Developer, Stakeholder • Analyst, Developer, Stakeholder, Leadership 1 2 3 4 5 6 7 Who does it? • Agile and may undergo iteration • Changes in Strategic goals, newer initiatives, releases, discoveries, reorgs • Sourcing/Blending challenges: Data handovers between systems, blending challenges • Scalability/automation • Data movements/latencies/ teams/approvals • Evolution of hypotheses, data changes/errors, success criteria • Competing priorities, data movements, Scenario Simulations • Success criteria, integration with research/testing tools, iterations • Integration with host systems, engineering investment, model tweaking, monitoring, customization Key Challenges
  • 11.
    COMPLICATION 3: OUTPUTOF ONE CAN BE INPUT/ADDITION TO ANOTHER Intended for Knowledge Sharing only 11 Behavioral Merchant Performance Clickstream/ Ops Campaign Performance VOC/Social/ CRM • Probability of Engagement/LTV Growth/Churn/Loyalty • Life event changes • Product/Price Migrations • Probability of Growth/Churn • Next Best Product/Offer • Network partners • Conversion Rate Optimization • Server Response Times • Time to Purchase • Campaign Responses • Next Best Product/Offer • Cross Channel target • Promoter/Detractor & drivers • Brand Appeal • Theme/entity of engagement Data Lake: Enriched with predictions e.g., Uber’s cross sell platform, Google Calendar, VDP
  • 12.
    COMPLICATION 4: REALDECISION MAKING NEEDS ADDITIONAL REASONING BEYOND ANALYTICS Intended for Knowledge Sharing only 12 Analytics provides insights into “actions”, Research context on “motivations” & Testing helps verify the “tactics” in the field and everything has to be productized… Strategy Data Tagging Data Platform Reporting Analytics Research Data Products Iterative Loop Why such complexity? Focus on Big Wins Reduced Wastage Quick Fixes Adaptability Assured execution Learning for future initiatives Optimization
  • 13.
    Intended for KnowledgeSharing only Intended for Knowledge Sharing only COMPLICATION 5: DEMANDS ON PREDICTIVE ANALYTICS HAVE INCREASED 13 Predictive Analytics Behavioral Analytics What are the customers doing? Voice of Customer What are the customers telling you? Platform Performance How are you delivering? Competitive Are the customers buying elsewhere? Social Listening How are customers discussing you? …aaanddd Better, Faster, Cheaper, Monetizable
  • 14.
    Intended for KnowledgeSharing only Quick recap of what it is Intended for Knowledge Sharing only So, what do we need then? 14
  • 15.
    Intended for KnowledgeSharing only • Extensible • Scalable • Flexible • Easy to integrate with other techniques Intended for Knowledge Sharing only HIGH LEVEL SUMMARY OF NEEDS: MODULAR, SHAREABLE & MONETIZABLE 15 keywordsuggest.org Iconfinder WebPT • Documentation • Governance • Integration with project management tools (collaboration) • Security/Privacy Management • Value Abstraction • API-able Modular Shareable Monetizable
  • 16.
    Intended for KnowledgeSharing only Quick recap of what it is Intended for Knowledge Sharing only Potential Solutions 16
  • 17.
    TWO DEPLOYMENT SOLUTIONS-PMML & PFA Intended for Knowledge Sharing only 17 Data Mining Group an independent Vendor Led Consortium that develops Data Mining Standards has come up with PMML (Predictive Model Mark Up Language) and PFA (Portable Format for Analytics) http://www.kdnuggets.com/2016/01/portable-format-analytics-models-production.html http://dmg.org/ https://www.ibm.com/developerworks/library/ba-predictive-analytics4/ba-predictive-analytics4-pdf.pdf https://www.ibm.com/developerworks/library/ba-ind-PMML1/ http://www.kdnuggets.com/faq/pmml.html https://journal.r-project.org/archive/2009-1/RJournal_2009-1_Guazzelli+et+al.pdf PMML PFA File XML JSON & YAML Maturity Mature but expanding Evolving Nesting/Customization Model Parameters Control Structures (Type System of Model Parameters & data - Callback function allowed) Flexibility Standard across most scoring engines (better than custom code) More flexible than PMML but safer than Custom Code Scope Data prep, Modeling, Scoring, Sharing +Pre/Post processing, enforced memory model
  • 18.
    PMML PROJECTS Intended forKnowledge Sharing only 18 http://data-informed.com/pmml-puts-big-data-to-work/
  • 19.
    POSITIONING OF PFA Intendedfor Knowledge Sharing only 19 http://data-informed.com/pmml-puts-big-data-to-work/
  • 20.
    Intended for KnowledgeSharing only Quick recap of what it is Intended for Knowledge Sharing only Why this, Why now, why here? 20
  • 21.
    Intended for KnowledgeSharing only Intended for Knowledge Sharing only BIGGER TRENDS THAT ARE SHAKING UP THE ANALYTICS WORLD FROM INSIDE OUT… 21 Demand Pressures: Complexity and nature of problems and their solutions, type of audience & consumption framework evolving Monetization opportunities- Direct, Indirect, Recurring Artificial Intelligence, IoE and “Smart”ening of devices/systems faster than expected. Evolution of input data sources and integration of multiple insights sources into decision making (A/B Testing, Research, Predictions/Scores from other models) Evolution from Service to Product to Platform (Build Once, Use Everywhere) …APIs are eating up our world
  • 22.
    Intended for KnowledgeSharing only Quick recap of what it is Intended for Knowledge Sharing only The parting words… 22
  • 23.
    SUMMARY Intended for KnowledgeSharing only Predictive Analytics has stopped being “one-off competitive edge project exercise” – it’s a necessary survival initiative for organizations Scale, complexity, breadth of needs (including Monetization) demand Platform approach. “Build Once, Use Everywhere” -consumption of predictive analytics outputs need to be easy to use, integrate, re-use/collaborate across multiple initiatives As everything becomes Productized via APIs, together they can become a business problem solving ANI 23 Streaming Analytics is quickly evolving into Streaming Predictive Analytics
  • 24.
    Intended for KnowledgeSharing only Quick recap of what it is Intended for Knowledge Sharing only Appendix
  • 25.
    THANK YOU! Intended forKnowledge Sharing only Would love to hear from you on any of the following forums… https://twitter.com/decisions_2_0 http://www.slideshare.net/RamkumarRavichandran https://www.youtube.com/channel/UCODSVC0WQws607clv0k8mQA/videos http://www.odbms.org/2015/01/ramkumar-ravichandran-visa/ https://www.linkedin.com/pub/ramkumar-ravichandran/10/545/67a RAMKUMAR RAVICHANDRAN
  • 26.
    Intended for KnowledgeSharing only Disclaimer: Participation is purely on a personal basis and does not represent VISA,Inc. in any form or matter. The talk is based on learning from work across industries and firms. Care has been taken to ensure no proprietary or work related info of any firm is used in any material. Director, Insights at Visa, Inc. Enable Decision Making at the Executives/ Product/Marketing level via actionable insights derived from Data. RAMKUMAR RAVICHANDRAN