Presentatie tmnl

Transcript

1. Build high quality
   ML models quickly
   using a central Feature Generator Library
   2023-09-12
   

   View Slide

2. Who are we?
   Nischay Ghattamaraju Roel Bertens
   

   View Slide

3. The Problem(s)
   What is holding teams back?
   

   View Slide

4. Solution
   Feature Generator Library (FGL)
   Collaboration
   § Quality
   § Single source of truth
   Reusability
   § Iteration / development speed
   § Consistency PoC and PROD
   § Re-usable code + documentation
   Efficient computation
   A python package
   containing well defined, reusable and tested features
   Managed dependencies
   between features
   Automated generation
   of documentation and diagrams
   What is it? Benefits
   

   View Slide

5. Comparison
   FGL vs Feature Store
   FGL Feature Store
   Stores the logic to generate features Stores the features
   Computation on demand Frequently triggered updates
   Computes only what is needed Precomputes all the features
   Retrieval is slow Retrieval is fast
   Easier to introduce Increases complexity of the platform
   No storage costs involved Storage costs are needed
   

   View Slide

6. Code Walkthroughs
   

   View Slide

7. Code Walkthroughs
   How we define a Feature
   

   View Slide

8. Code Walkthroughs
   How we define
   a Feature Group
   

   View Slide

9. Code Walkthroughs
   Example Feature Group
   

   View Slide

10. Code Walkthroughs
    Usage of the FGL
    Client_id avg_nr_of_items avg_nr_of_items__
    max_per_city
    avg_nr_of_items__
    compared_to_city_max
    city
    client_1 3 3 1.0 Utrecht
    client_2 2 3 0.66 Utrecht
    

    View Slide

11. Learnings
    Some of the lessons we learnt
    Thorough documentation makes it easier to collaborate
    Invest if you want to scale to more features/models
    (so no custom feature pipelines)
    Don’t start with a Feature Store but start simple and upgrade when needed
    Think about who is going to own and maintain the code.
    Communicate the value clearly to users and contributors to avoid shelfware
    

    View Slide

12. Questions?
    

    View Slide