Identifying And Prototyping Data Science Use Cases

Identifying And Prototyping Data Science Use Cases

• 1.
  Identifying and Prototyping Data Science Use Cases How to build an organizational capability and set it up for success, not failure Ambrus Vancso
• 2.
  60%
• 3.
  The Problem Data projects fail too often page 03*Nick Heudecker is VP Analyst at Gartner *
• 4.
  Where do Data Science Projects go Wrong page 04 No Clear Vision On Business Goal Models Not Understood By Stakeholders Lack Of Sponsorship And Other Project Woes False Assumptions About Data Results Never Deployed Not The Right Skill Mix On Team Challenges
• 5.
  Identify
• 6.
  page 06 EVALUATE Clarify Understand Qualify Identify Impact (cost/revenue/capability) Define Metric PRIORITIZE Value Effort Need Visualize Communicate CAPTURE Involve Key Stakeholders Prepare Questionnaire / Interview / Workshop Record Organisation Business Process Persona Data Systems Use Case Identification Process Define Value for Organization
• 7.
  Capture Key Questions to Understand page 07
• 8.
  Capture Example One Pager page 08
• 9.
  Evaluate Clarify if BI or DS page 09 Is it a Data Science Project at all?
• 10.
  Prioritize Joint Decision page 010 Effort Value
• 11.
  The Cycle
• 12.
  The Scientific Method As an Ongoing Process page 012 Observe Develop General Theory Build Hypothesis Testable Predictions Ask questionsGather Data to Test Predictions Iterate
• 13.
  Agile Process As in Software Engineering page 013 Multidisciplinary team Iteration Increased Understanding and Focuson Delivery of Real Value Product Backlog Release Backlog Sprint Backlog Product Release Retrospective
• 14.
  KNOWN UNKNOWN Types of Data Science Projects Engineering vs Science Paradigm page 014 Question DataKNOWN UNKNOWN R&DExploration DiscoveryIteration Addressable space R&D
• 15.
  An Agile Process Best of Both Worlds page 015 PrototypeUse Case 03 02 04 01 DELIVER tangible result in every sprint, aim for end-to-end as much as possible TIMEBOXING and continuous backlog revision is the key to bridge approaches DEFINITION OF DONE is essential and not native to most Data Scientists ITERATION is the most common denominator between scientific and agile engineering paradigm
• 16.
  Prototype
• 17.
  What is a Prototype? Shared Understanding is Key page 017 POH or POC Proof of Hypothesis, can be part of a POC An experiment basically Something tangible coming out of the project that has the potential of practical usage and scaling Prototype Something ready for production deployment without significant rework MVP Definition of done
• 18.
  Use a Prototype to Understand and Validate page 018 Data Feasibility TeamPlatformNeed Value
• 19.
  page 019 Skills and Roles
• 20.
  Big Enough vs Big Data Progress is much, much faster and easier if tools and data fits into memory page 020 VS. Note: you can rent machines of up to TiB range on the cloud. GPU instances also available. A memory optimized node of 488 GiBis <5 $/hr.
• 21.
  Data Handling Paradigms Optimize for Persona - can also Mix & Match page 021 Code Powerful, Open, Difficult Expert Data Scientist, Developer Engineering friendly Jupyter, RStudio Pipeline Powerful, Specific, Learnable Citizen Data Scientist, Analyst Middle ground, “Lingua Franca” Alteryx, KNIME, Rapidminer Table Limited, Narrow, Intuitive Analyst, Office User Everybody knows spreadsheets Excel -> Power BI
• 22.
  Be Ready To Explain Always Have a White Box Model (As Well) page 022 VS.
• 23.
  Communicate, Visually Tell the story – in light of your key Personae page 023 0 200 400 600 800 1 000 1200 1 400 1 600 jan feb mar apr may jun jul aug sep oct nov dec Approach One Approach Two Approach Three
• 24.
  Operationalize
• 25.
  Operationalize and Scale Infrastructure, Process, Organizational Challenges page 025* Source: Hidden Technical Debt in Machine Learning Systems https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning- systems.pdhttp:/martin.zinkevich.org/rules_of_ml/rules_of_ml ML Code
• 26.
  Operationalize and Scale Data Science Platforms Can Give a Head Start page 026 Open Source Fully open source, open core or business source offerings. Gartner 2018 Magic Quadrant for Data Science and Machine Learning Platforms Source: https://www.kdnuggets.com/2018/02/gartner-2018-mq-data-science-machine-learning-changes.html Note: Gartner research methodology does not include open source platforms like R and Python. While these are powerful, a platform shines in its comprehensiveness rather, especially when it comes to ease of deployment. Also most open platforms already allow to run R/Py scripts and incorporate the results.
• 27.
  Thank you @ambrusvancso linkedin.com/in/ambrusvancso/ +44 7937 962 169