Demystifying ML/AI

Decoding ML/AI
A Practical Guide For Startups To
Drive Growth And Innovation
Denver Startup Week, 2023
PRESENTING SPONSOR
TITLE SPONSORS
TRACK SPONSORS
HEADLINE SPONSORS
PARTNER SPONSORS MEMBER SPONSORS
- CBRE
- Colorado Sun
- The Commons on Champa
- DAT Software and Analytics
- Denver Place
- Expansive Workspace
- Greenspoon Marder
- Halliburton Labs
- Jake Jabs Center for Entrepreneurship
- Molson Coors
- MSU Denver
- Park Central
- Polsinelli
- Tea with Tae
- Bounce Back
- Caruso
- Credera
- Doyle Group
- Industrious
- Initial Capacity
- MLJ Insurance
- Nexton
- Spectrum
- WeWork
© 2022 Ibotta, Inc. Proprietary and confidential, not to be shared without Ibotta’s express consent.
2.0
addition
Argie Angeleas
Group Product Manager
- @ibotta
Taylor Names
Principal ML Engineer
- @ibotta
Matt Reynolds
Principal ML Platform
Engineer - @ibotta
Gartner Predicts That 85%
Of ML Projects Will Fail
Demystifying ML/AI
Denver, CO - 2023
Edtech Startup
People convert to registered users, BUT
they don’t watch courses
Course CTR (Click-through-rate) after a
strong start stayed flat
© 2023 Ibotta, Inc. Proprietary and confidential, not to be shared without Ibotta’s express consent.
Product
Where To Start?
3 Important Questions
What is the problem we
are trying to solve?
Are there business
constraints?
Understand The Problem..
Remove Ambiguity..
But How?
What does a successful
outcome like?
Let The Fun Begin Step.0
Define The Problem Spaces
People can’t find relevant coursework and therefore they
are not starting courses
People find relevant courses, but the content of each
course does not meet their expectations
People can’t understand how to use our website, and
that’s why they are not starting courses
Great, But What’s Next?
Peel the onion
Is this an ML problem? What is the project value?
What is the expectation
in business?
Let The Fun Begin Vol. 1
Define The Hypotheses
Secret #1: Don’t Do It In A Dark Room On Your Own
Involve Subject Matter Experts
(Business, Engineering, Data Science, Product)
Define hypotheses that can solve the problem
ML is different than traditional software - it is
hypotheses based
© 2023 Ibotta, Inc. Proprietary and confidential, not to be shared without Ibotta’s express consent.
Let the fun begin vol. 2
Ideate solutions
Earn trust through outcomes
Let The Fun Begin Vol. 3
Define Your Metrics
Secret #2: Don’t do it in a dark room on your own
Success Metric:
Click through rate
Guardrail Metrics:
› First course completion rate
› Sponsored content click through rate
© 2023 Ibotta, Inc. Proprietary and confidential, not to be shared without Ibotta’s express consent.
Data Science
Your First Model Will Fail
Before model experimentation starts, there is uncertainty around how effectively the business
problem can be solved
› Should we expect a 2% lift in click through rate or 20%?
Uncertainty around which data will be predictive
› Example hypothesis: student’s geographic location will help predict relevant courses
› Many hypotheses won’t pan out!
Model experimentation process should be optimized for rapid iteration
› Input data + modeling approach = success?
› Don’t over polish a solution that isn’t guaranteed to succeed
How Will I Know If My Model Works?
In A Perfect World, Get Model Predictions In Front Of Real Customers & Measure Business
Impact!
A/B Testing: split customers between treatment (ML powered) and control (status-quo)
Cons:
› Large engineering investment (reliable production-ready solution)
› Risk to business (can we afford to see a significant drop in business metrics in our test population?)
› Slow and difficult to get conclusive results (if you have a small user base)
Is there a lower cost and lower risk alternative for rapid model evaluation?
Yes! Evaluate model performance on historical data
› Commonly referred to as offline testing or backtesting
Estimating Performance On Historical Data
Can you directly measure the business impact metric with different model
approaches on past data?
› If yes, do this!
01
What if it is infeasible to measure business metric on historical data?
› Are there proxy metrics that are well correlated with the critical business impact metrics?
• Example: Click through rate -> how high did the model rank items that were clicked
02
Remember to compare performance against a reasonable heuristic!
› Does my model generate more clicks than content sorted in order of general popularity?
03
Don’t over-optimize “offline” performance. Get in front of real users!
04
Clean, representative, and predictive input (training) data
› Clean: outliers, erroneous data need to be handled with care
› Representative: are the business scenarios where the model
will add value represented in your training set?
› Predictive: talk with SMEs on what data should be sourced
and transformed
The right model for the right problem
Rules of thumb:
› Tabular data: ensemble decision trees (XGBoost, lightGBM,
random forest)
› Text/computer vision/audio: open source pre-trained
neural networks
Model training is less than 10% of the total project effort!
What Does It
Take To Build An
Accurate Model?
Is My Model Ready For The Big Leagues?
Am I reasonably confident that my
model will add value & do no harm?
Outperforms a heuristic according to
business metric or reasonable proxy metric
Was my methodology of
evaluation sound?
Always peer review
code/analysis/conclusions!
Is my model able to be reasonably
supported in a production environment?
Overly complex models can be difficult to
productionize and expensive to maintain
If so, SHIP IT!
© 2023 Ibotta, Inc. Proprietary and confidential, not to be shared without Ibotta’s express consent. 22
Machine Learning Infrastructure
OK, I Have A Model, Now What?
You Need To Figure Out The Best Way To Get It Into Production
1
2
3
4
How “fresh” does the
model need to be?
What’s the simplest way to get
the results?
How often do you need to
perform predictions?
Who do you need to
create predictions for?
If you haven’t been talking to the engineers who will use the results, start now
Start with batch predictions
Starting With Batch Predictions Allows You To Avoid All This:
Real time feature store
(features = data used by model
to generate predictions)
Model serving code
Additional infrastructure
Latency concerns
Resiliency issues
01
02 04
03 05
25
25
› Use source control
› Have others review changes
› Be consistent – use naming conventions
› Parameterize for different environments/versions
› Provide a defined interface for your model predictions
› Reduce manual input in pipelines – aim for CI/CD
Software Engineering Teams (Should!) Do This, Copy Them
Follow Basic
Engineering Best
Practices
What Could This Look Like?
Source Control:
› Consider data scientist model work (generally Jupyter notebook)*
› Scripts for model data generation, training, batch predictions and storing results
Cloud storage:
Standardized naming model artifact, folder by environment, training date
Compute instance/Serverless:
Run scripts on regular time schedule or some automated trigger
Database:
Prediction results - include model version column
How can I save time?
Avoid scope creep
(stick to your defined
goals)
v
v
v
v
v
v
v
v
v
v
v
v
v
v
Keep it simple
(don’t try and build real-time
out the gate)
Use managed services
(don’t build everything from
scratch)
Get dedicated staff involved for
end to end support
(project should be
self-sufficient)
Follow best practices
(use standard model
approaches)
Use what you already have
(data engineering, service tools)
Avoid analysis paralysis
(beats baseline = start test)
Hire all-rounders
(no research only data
scientists)
What Should I Avoid Cutting Corners On?
Get other teams involved early
(upstream and downstream)
Don’t skip on automation
Regularly review
(both internal for design/code and stakeholders for progress)
Provide an API interface
(reduce coupling)
Basic monitoring
(did training fail? Did my predictions job run?)
Collaboration works wonders → business <> engineering <> data science <> product
Utilizing a structured approach to problem definition increases the likelihood of success
Even simple models can create impact that gets the business raving
Conclusions
Thank You
Questions?
1 de 30

Recomendados

Agile Mumbai 2022 - Ashwinee Singh | Agile in AI or AI in Agile? por
Agile Mumbai 2022 - Ashwinee Singh | Agile in AI or AI in Agile?Agile Mumbai 2022 - Ashwinee Singh | Agile in AI or AI in Agile?
Agile Mumbai 2022 - Ashwinee Singh | Agile in AI or AI in Agile?AgileNetwork
46 visualizações37 slides
SDD2017 - 03 Abed Ajraou - putting data science in your business a first uti... por
SDD2017 - 03 Abed Ajraou  - putting data science in your business a first uti...SDD2017 - 03 Abed Ajraou  - putting data science in your business a first uti...
SDD2017 - 03 Abed Ajraou - putting data science in your business a first uti...Dario Mangano
238 visualizações28 slides
Putting data science in your business a first utility feedback por
Putting data science in your business a first utility feedbackPutting data science in your business a first utility feedback
Putting data science in your business a first utility feedbackPeculium Crypto
520 visualizações28 slides
[DSC Europe 22] The Making of a Data Organization - Denys Holovatyi por
[DSC Europe 22] The Making of a Data Organization - Denys Holovatyi[DSC Europe 22] The Making of a Data Organization - Denys Holovatyi
[DSC Europe 22] The Making of a Data Organization - Denys HolovatyiDataScienceConferenc1
17 visualizações16 slides
BDW17 London - Abed Ajraou - First Utility - Putting Data Science in your Bus... por
BDW17 London - Abed Ajraou - First Utility - Putting Data Science in your Bus...BDW17 London - Abed Ajraou - First Utility - Putting Data Science in your Bus...
BDW17 London - Abed Ajraou - First Utility - Putting Data Science in your Bus...Big Data Week
70 visualizações25 slides
Maximising likelihood of success: Applying Product Management to AI/ML/DS pr... por
Maximising likelihood of success:  Applying Product Management to AI/ML/DS pr...Maximising likelihood of success:  Applying Product Management to AI/ML/DS pr...
Maximising likelihood of success: Applying Product Management to AI/ML/DS pr...Kevin Wong
139 visualizações19 slides

Mais conteúdo relacionado

Similar a Demystifying ML/AI

Artificial intelligence capabilities overview yashowardhan sowale cwin18-india por
Artificial intelligence capabilities overview yashowardhan sowale cwin18-indiaArtificial intelligence capabilities overview yashowardhan sowale cwin18-india
Artificial intelligence capabilities overview yashowardhan sowale cwin18-indiaCapgemini
242 visualizações19 slides
IoT Product Design and Prototyping por
IoT Product Design and PrototypingIoT Product Design and Prototyping
IoT Product Design and PrototypingDr. Shivananda Koteshwar
671 visualizações43 slides
Top Takeaways from Validate 2019 por
Top Takeaways from Validate 2019Top Takeaways from Validate 2019
Top Takeaways from Validate 2019ObservePoint
186 visualizações30 slides
Cloud Analytics Playbook por
Cloud Analytics PlaybookCloud Analytics Playbook
Cloud Analytics PlaybookBooz Allen Hamilton
3K visualizações36 slides
Enterprise Architecture: Part I - Contextualizing the Practice por
Enterprise Architecture: Part I - Contextualizing the PracticeEnterprise Architecture: Part I - Contextualizing the Practice
Enterprise Architecture: Part I - Contextualizing the PracticeFru Louis
9.3K visualizações16 slides
Pin the tail on the metric v00 75 min version por
Pin the tail on the metric v00 75 min versionPin the tail on the metric v00 75 min version
Pin the tail on the metric v00 75 min versionSteven Martin
237 visualizações49 slides

Similar a Demystifying ML/AI(20)

Artificial intelligence capabilities overview yashowardhan sowale cwin18-india por Capgemini
Artificial intelligence capabilities overview yashowardhan sowale cwin18-indiaArtificial intelligence capabilities overview yashowardhan sowale cwin18-india
Artificial intelligence capabilities overview yashowardhan sowale cwin18-india
Capgemini242 visualizações
Top Takeaways from Validate 2019 por ObservePoint
Top Takeaways from Validate 2019Top Takeaways from Validate 2019
Top Takeaways from Validate 2019
ObservePoint186 visualizações
Cloud Analytics Playbook por Booz Allen Hamilton
Cloud Analytics PlaybookCloud Analytics Playbook
Cloud Analytics Playbook
Booz Allen Hamilton3K visualizações
Enterprise Architecture: Part I - Contextualizing the Practice por Fru Louis
Enterprise Architecture: Part I - Contextualizing the PracticeEnterprise Architecture: Part I - Contextualizing the Practice
Enterprise Architecture: Part I - Contextualizing the Practice
Fru Louis9.3K visualizações
Pin the tail on the metric v00 75 min version por Steven Martin
Pin the tail on the metric v00 75 min versionPin the tail on the metric v00 75 min version
Pin the tail on the metric v00 75 min version
Steven Martin237 visualizações
Evolving Towards a Modernized Platform: Our Success Story por VMware Tanzu
Evolving Towards a Modernized Platform: Our Success StoryEvolving Towards a Modernized Platform: Our Success Story
Evolving Towards a Modernized Platform: Our Success Story
VMware Tanzu100 visualizações
Evolving Towards a Modernized Platform: Our Success Story por VMware Tanzu
Evolving Towards a Modernized Platform: Our Success StoryEvolving Towards a Modernized Platform: Our Success Story
Evolving Towards a Modernized Platform: Our Success Story
VMware Tanzu242 visualizações
Surge engr 245 lean launchpad stanford 2020 por Stanford University
Surge engr 245 lean launchpad stanford 2020Surge engr 245 lean launchpad stanford 2020
Surge engr 245 lean launchpad stanford 2020
Stanford University2.1K visualizações
Betsol | Machine Learning for IT Project Estimates por BETSOL
Betsol | Machine Learning for IT Project Estimates  Betsol | Machine Learning for IT Project Estimates
Betsol | Machine Learning for IT Project Estimates
BETSOL212 visualizações
Bridging the Gap: Analyzing Data in and Below the Cloud por Inside Analysis
Bridging the Gap: Analyzing Data in and Below the CloudBridging the Gap: Analyzing Data in and Below the Cloud
Bridging the Gap: Analyzing Data in and Below the Cloud
Inside Analysis644 visualizações
How to Build an AI/ML Product and Sell it by SalesChoice CPO por Product School
How to Build an AI/ML Product and Sell it by SalesChoice CPOHow to Build an AI/ML Product and Sell it by SalesChoice CPO
How to Build an AI/ML Product and Sell it by SalesChoice CPO
Product School489 visualizações
Tdwi march 2015 presentation por Alison Macfie
Tdwi march 2015 presentationTdwi march 2015 presentation
Tdwi march 2015 presentation
Alison Macfie457 visualizações
Using Data Science to Build an End-to-End Recommendation System por VMware Tanzu
Using Data Science to Build an End-to-End Recommendation SystemUsing Data Science to Build an End-to-End Recommendation System
Using Data Science to Build an End-to-End Recommendation System
VMware Tanzu1.6K visualizações
Architecting a Data Platform For Enterprise Use (Strata NY 2018) por mark madsen
Architecting a Data Platform For Enterprise Use (Strata NY 2018)Architecting a Data Platform For Enterprise Use (Strata NY 2018)
Architecting a Data Platform For Enterprise Use (Strata NY 2018)
mark madsen1.1K visualizações
Leverage Data Strategy as a Catalyst for Innovation por Glorium Tech
Leverage Data Strategy as a Catalyst for InnovationLeverage Data Strategy as a Catalyst for Innovation
Leverage Data Strategy as a Catalyst for Innovation
Glorium Tech38 visualizações
AI Orange Belt - Session 3 por AI Black Belt
AI Orange Belt - Session 3AI Orange Belt - Session 3
AI Orange Belt - Session 3
AI Black Belt1.3K visualizações
Intro to Data Analytics with Oscar's Director of Product por Product School
 Intro to Data Analytics with Oscar's Director of Product Intro to Data Analytics with Oscar's Director of Product
Intro to Data Analytics with Oscar's Director of Product
Product School878 visualizações

Último

iSAQB Software Architecture Gathering 2023: How Process Orchestration Increas... por
iSAQB Software Architecture Gathering 2023: How Process Orchestration Increas...iSAQB Software Architecture Gathering 2023: How Process Orchestration Increas...
iSAQB Software Architecture Gathering 2023: How Process Orchestration Increas...Bernd Ruecker
50 visualizações69 slides
DRBD Deep Dive - Philipp Reisner - LINBIT por
DRBD Deep Dive - Philipp Reisner - LINBITDRBD Deep Dive - Philipp Reisner - LINBIT
DRBD Deep Dive - Philipp Reisner - LINBITShapeBlue
62 visualizações21 slides
Uni Systems for Power Platform.pptx por
Uni Systems for Power Platform.pptxUni Systems for Power Platform.pptx
Uni Systems for Power Platform.pptxUni Systems S.M.S.A.
58 visualizações21 slides
CloudStack Object Storage - An Introduction - Vladimir Petrov - ShapeBlue por
CloudStack Object Storage - An Introduction - Vladimir Petrov - ShapeBlueCloudStack Object Storage - An Introduction - Vladimir Petrov - ShapeBlue
CloudStack Object Storage - An Introduction - Vladimir Petrov - ShapeBlueShapeBlue
46 visualizações15 slides
Elevating Privacy and Security in CloudStack - Boris Stoyanov - ShapeBlue por
Elevating Privacy and Security in CloudStack - Boris Stoyanov - ShapeBlueElevating Privacy and Security in CloudStack - Boris Stoyanov - ShapeBlue
Elevating Privacy and Security in CloudStack - Boris Stoyanov - ShapeBlueShapeBlue
96 visualizações7 slides
Hypervisor Agnostic DRS in CloudStack - Brief overview & demo - Vishesh Jinda... por
Hypervisor Agnostic DRS in CloudStack - Brief overview & demo - Vishesh Jinda...Hypervisor Agnostic DRS in CloudStack - Brief overview & demo - Vishesh Jinda...
Hypervisor Agnostic DRS in CloudStack - Brief overview & demo - Vishesh Jinda...ShapeBlue
63 visualizações13 slides

Último(20)

iSAQB Software Architecture Gathering 2023: How Process Orchestration Increas... por Bernd Ruecker
iSAQB Software Architecture Gathering 2023: How Process Orchestration Increas...iSAQB Software Architecture Gathering 2023: How Process Orchestration Increas...
iSAQB Software Architecture Gathering 2023: How Process Orchestration Increas...
Bernd Ruecker50 visualizações
DRBD Deep Dive - Philipp Reisner - LINBIT por ShapeBlue
DRBD Deep Dive - Philipp Reisner - LINBITDRBD Deep Dive - Philipp Reisner - LINBIT
DRBD Deep Dive - Philipp Reisner - LINBIT
ShapeBlue62 visualizações
Uni Systems for Power Platform.pptx por Uni Systems S.M.S.A.
Uni Systems for Power Platform.pptxUni Systems for Power Platform.pptx
Uni Systems for Power Platform.pptx
Uni Systems S.M.S.A.58 visualizações
CloudStack Object Storage - An Introduction - Vladimir Petrov - ShapeBlue por ShapeBlue
CloudStack Object Storage - An Introduction - Vladimir Petrov - ShapeBlueCloudStack Object Storage - An Introduction - Vladimir Petrov - ShapeBlue
CloudStack Object Storage - An Introduction - Vladimir Petrov - ShapeBlue
ShapeBlue46 visualizações
Elevating Privacy and Security in CloudStack - Boris Stoyanov - ShapeBlue por ShapeBlue
Elevating Privacy and Security in CloudStack - Boris Stoyanov - ShapeBlueElevating Privacy and Security in CloudStack - Boris Stoyanov - ShapeBlue
Elevating Privacy and Security in CloudStack - Boris Stoyanov - ShapeBlue
ShapeBlue96 visualizações
Hypervisor Agnostic DRS in CloudStack - Brief overview & demo - Vishesh Jinda... por ShapeBlue
Hypervisor Agnostic DRS in CloudStack - Brief overview & demo - Vishesh Jinda...Hypervisor Agnostic DRS in CloudStack - Brief overview & demo - Vishesh Jinda...
Hypervisor Agnostic DRS in CloudStack - Brief overview & demo - Vishesh Jinda...
ShapeBlue63 visualizações
GDG Cloud Southlake 28 Brad Taylor and Shawn Augenstein Old Problems in the N... por James Anderson
GDG Cloud Southlake 28 Brad Taylor and Shawn Augenstein Old Problems in the N...GDG Cloud Southlake 28 Brad Taylor and Shawn Augenstein Old Problems in the N...
GDG Cloud Southlake 28 Brad Taylor and Shawn Augenstein Old Problems in the N...
James Anderson133 visualizações
MVP and prioritization.pdf por rahuldharwal141
MVP and prioritization.pdfMVP and prioritization.pdf
MVP and prioritization.pdf
rahuldharwal14138 visualizações
Network Source of Truth and Infrastructure as Code revisited por Network Automation Forum
Network Source of Truth and Infrastructure as Code revisitedNetwork Source of Truth and Infrastructure as Code revisited
Network Source of Truth and Infrastructure as Code revisited
Network Automation Forum42 visualizações
Microsoft Power Platform.pptx por Uni Systems S.M.S.A.
Microsoft Power Platform.pptxMicrosoft Power Platform.pptx
Microsoft Power Platform.pptx
Uni Systems S.M.S.A.67 visualizações
Transitioning from VMware vCloud to Apache CloudStack: A Path to Profitabilit... por ShapeBlue
Transitioning from VMware vCloud to Apache CloudStack: A Path to Profitabilit...Transitioning from VMware vCloud to Apache CloudStack: A Path to Profitabilit...
Transitioning from VMware vCloud to Apache CloudStack: A Path to Profitabilit...
ShapeBlue57 visualizações
Migrating VMware Infra to KVM Using CloudStack - Nicolas Vazquez - ShapeBlue por ShapeBlue
Migrating VMware Infra to KVM Using CloudStack - Nicolas Vazquez - ShapeBlueMigrating VMware Infra to KVM Using CloudStack - Nicolas Vazquez - ShapeBlue
Migrating VMware Infra to KVM Using CloudStack - Nicolas Vazquez - ShapeBlue
ShapeBlue96 visualizações
Five Things You SHOULD Know About Postman por Postman
Five Things You SHOULD Know About PostmanFive Things You SHOULD Know About Postman
Five Things You SHOULD Know About Postman
Postman40 visualizações
Ransomware is Knocking your Door_Final.pdf por Security Bootcamp
Ransomware is Knocking your Door_Final.pdfRansomware is Knocking your Door_Final.pdf
Ransomware is Knocking your Door_Final.pdf
Security Bootcamp76 visualizações
NTGapps NTG LowCode Platform por Mustafa Kuğu
NTGapps NTG LowCode Platform NTGapps NTG LowCode Platform
NTGapps NTG LowCode Platform
Mustafa Kuğu141 visualizações
Kyo - Functional Scala 2023.pdf por Flavio W. Brasil
Kyo - Functional Scala 2023.pdfKyo - Functional Scala 2023.pdf
Kyo - Functional Scala 2023.pdf
Flavio W. Brasil434 visualizações
What’s New in CloudStack 4.19 - Abhishek Kumar - ShapeBlue por ShapeBlue
What’s New in CloudStack 4.19 - Abhishek Kumar - ShapeBlueWhat’s New in CloudStack 4.19 - Abhishek Kumar - ShapeBlue
What’s New in CloudStack 4.19 - Abhishek Kumar - ShapeBlue
ShapeBlue131 visualizações
VNF Integration and Support in CloudStack - Wei Zhou - ShapeBlue por ShapeBlue
VNF Integration and Support in CloudStack - Wei Zhou - ShapeBlueVNF Integration and Support in CloudStack - Wei Zhou - ShapeBlue
VNF Integration and Support in CloudStack - Wei Zhou - ShapeBlue
ShapeBlue85 visualizações
HTTP headers that make your website go faster - devs.gent November 2023 por Thijs Feryn
HTTP headers that make your website go faster - devs.gent November 2023HTTP headers that make your website go faster - devs.gent November 2023
HTTP headers that make your website go faster - devs.gent November 2023
Thijs Feryn28 visualizações

Demystifying ML/AI

  • 1. Decoding ML/AI A Practical Guide For Startups To Drive Growth And Innovation Denver Startup Week, 2023
  • 3. HEADLINE SPONSORS PARTNER SPONSORS MEMBER SPONSORS - CBRE - Colorado Sun - The Commons on Champa - DAT Software and Analytics - Denver Place - Expansive Workspace - Greenspoon Marder - Halliburton Labs - Jake Jabs Center for Entrepreneurship - Molson Coors - MSU Denver - Park Central - Polsinelli - Tea with Tae - Bounce Back - Caruso - Credera - Doyle Group - Industrious - Initial Capacity - MLJ Insurance - Nexton - Spectrum - WeWork
  • 4. © 2022 Ibotta, Inc. Proprietary and confidential, not to be shared without Ibotta’s express consent. 2.0 addition Argie Angeleas Group Product Manager - @ibotta Taylor Names Principal ML Engineer - @ibotta Matt Reynolds Principal ML Platform Engineer - @ibotta
  • 5. Gartner Predicts That 85% Of ML Projects Will Fail
  • 8. Edtech Startup People convert to registered users, BUT they don’t watch courses Course CTR (Click-through-rate) after a strong start stayed flat
  • 9. © 2023 Ibotta, Inc. Proprietary and confidential, not to be shared without Ibotta’s express consent. Product
  • 10. Where To Start? 3 Important Questions What is the problem we are trying to solve? Are there business constraints? Understand The Problem.. Remove Ambiguity.. But How? What does a successful outcome like?
  • 11. Let The Fun Begin Step.0 Define The Problem Spaces People can’t find relevant coursework and therefore they are not starting courses People find relevant courses, but the content of each course does not meet their expectations People can’t understand how to use our website, and that’s why they are not starting courses
  • 12. Great, But What’s Next? Peel the onion Is this an ML problem? What is the project value? What is the expectation in business?
  • 13. Let The Fun Begin Vol. 1 Define The Hypotheses Secret #1: Don’t Do It In A Dark Room On Your Own Involve Subject Matter Experts (Business, Engineering, Data Science, Product) Define hypotheses that can solve the problem ML is different than traditional software - it is hypotheses based
  • 14. © 2023 Ibotta, Inc. Proprietary and confidential, not to be shared without Ibotta’s express consent. Let the fun begin vol. 2 Ideate solutions Earn trust through outcomes
  • 15. Let The Fun Begin Vol. 3 Define Your Metrics Secret #2: Don’t do it in a dark room on your own Success Metric: Click through rate Guardrail Metrics: › First course completion rate › Sponsored content click through rate
  • 16. © 2023 Ibotta, Inc. Proprietary and confidential, not to be shared without Ibotta’s express consent. Data Science
  • 17. Your First Model Will Fail Before model experimentation starts, there is uncertainty around how effectively the business problem can be solved › Should we expect a 2% lift in click through rate or 20%? Uncertainty around which data will be predictive › Example hypothesis: student’s geographic location will help predict relevant courses › Many hypotheses won’t pan out! Model experimentation process should be optimized for rapid iteration › Input data + modeling approach = success? › Don’t over polish a solution that isn’t guaranteed to succeed
  • 18. How Will I Know If My Model Works? In A Perfect World, Get Model Predictions In Front Of Real Customers & Measure Business Impact! A/B Testing: split customers between treatment (ML powered) and control (status-quo) Cons: › Large engineering investment (reliable production-ready solution) › Risk to business (can we afford to see a significant drop in business metrics in our test population?) › Slow and difficult to get conclusive results (if you have a small user base) Is there a lower cost and lower risk alternative for rapid model evaluation? Yes! Evaluate model performance on historical data › Commonly referred to as offline testing or backtesting
  • 19. Estimating Performance On Historical Data Can you directly measure the business impact metric with different model approaches on past data? › If yes, do this! 01 What if it is infeasible to measure business metric on historical data? › Are there proxy metrics that are well correlated with the critical business impact metrics? • Example: Click through rate -> how high did the model rank items that were clicked 02 Remember to compare performance against a reasonable heuristic! › Does my model generate more clicks than content sorted in order of general popularity? 03 Don’t over-optimize “offline” performance. Get in front of real users! 04
  • 20. Clean, representative, and predictive input (training) data › Clean: outliers, erroneous data need to be handled with care › Representative: are the business scenarios where the model will add value represented in your training set? › Predictive: talk with SMEs on what data should be sourced and transformed The right model for the right problem Rules of thumb: › Tabular data: ensemble decision trees (XGBoost, lightGBM, random forest) › Text/computer vision/audio: open source pre-trained neural networks Model training is less than 10% of the total project effort! What Does It Take To Build An Accurate Model?
  • 21. Is My Model Ready For The Big Leagues? Am I reasonably confident that my model will add value & do no harm? Outperforms a heuristic according to business metric or reasonable proxy metric Was my methodology of evaluation sound? Always peer review code/analysis/conclusions! Is my model able to be reasonably supported in a production environment? Overly complex models can be difficult to productionize and expensive to maintain If so, SHIP IT!
  • 22. © 2023 Ibotta, Inc. Proprietary and confidential, not to be shared without Ibotta’s express consent. 22 Machine Learning Infrastructure
  • 23. OK, I Have A Model, Now What? You Need To Figure Out The Best Way To Get It Into Production 1 2 3 4 How “fresh” does the model need to be? What’s the simplest way to get the results? How often do you need to perform predictions? Who do you need to create predictions for? If you haven’t been talking to the engineers who will use the results, start now
  • 24. Start with batch predictions Starting With Batch Predictions Allows You To Avoid All This: Real time feature store (features = data used by model to generate predictions) Model serving code Additional infrastructure Latency concerns Resiliency issues 01 02 04 03 05
  • 25. 25 25 › Use source control › Have others review changes › Be consistent – use naming conventions › Parameterize for different environments/versions › Provide a defined interface for your model predictions › Reduce manual input in pipelines – aim for CI/CD Software Engineering Teams (Should!) Do This, Copy Them Follow Basic Engineering Best Practices
  • 26. What Could This Look Like? Source Control: › Consider data scientist model work (generally Jupyter notebook)* › Scripts for model data generation, training, batch predictions and storing results Cloud storage: Standardized naming model artifact, folder by environment, training date Compute instance/Serverless: Run scripts on regular time schedule or some automated trigger Database: Prediction results - include model version column
  • 27. How can I save time? Avoid scope creep (stick to your defined goals) v v v v v v v v v v v v v v Keep it simple (don’t try and build real-time out the gate) Use managed services (don’t build everything from scratch) Get dedicated staff involved for end to end support (project should be self-sufficient) Follow best practices (use standard model approaches) Use what you already have (data engineering, service tools) Avoid analysis paralysis (beats baseline = start test) Hire all-rounders (no research only data scientists)
  • 28. What Should I Avoid Cutting Corners On? Get other teams involved early (upstream and downstream) Don’t skip on automation Regularly review (both internal for design/code and stakeholders for progress) Provide an API interface (reduce coupling) Basic monitoring (did training fail? Did my predictions job run?)
  • 29. Collaboration works wonders → business <> engineering <> data science <> product Utilizing a structured approach to problem definition increases the likelihood of success Even simple models can create impact that gets the business raving Conclusions