Scaling Machine Learning at Holiday Extras (Big Data LDN 2019))

Transcript

1. Scaling Machine Learning at Holiday Extras REBECCA VICKERY | DATA SCIENTIST @vickdata
2. Travel planning is time consuming Airport parking Airport hotels Airport lounges Travel insurance Holiday money Port products Car hire Airport transfers 582 minutes Over 46 days* Travel Planning *Facebook commissioned consumer research company GfK
3. Optimising consumer decision making Airport parking Airport hotels Airport lounges Travel insurance Holiday money Port products Car hire Airport transfers Less Hassle. More Holiday Trip recommendations
4. Automated bidding Ad targeting Channel optimisation 1 Ad spend 2 Commercial 3 Customer Experience 4 Marketing Lots of other processes to optimise Automated pricing Allocation Revenue optimisation Automated call handling Personalised experiences Intelligent messaging Optimise send frequency
5. How to scale Use Cases and Buy in (Input Team Deployment
6. How to scale Use Cases and Buy in (Input Team Deployment
7. “Ideas are worth nothing unless executed”, Derek Sivers
8. Deploying machine learning is hard Scaling is even harder
9. Tools - Data Scientists Open source Lack Software Development expertise Mainly Python c Flaticon
10. Tools - Software Engineers Different tools Lack ML/Data expertise Mainly Javascript c Flaticon
11. Data science process The wrong kind of independence c Flaticon
12. People Small data science team Science + software experts are rare c Flaticon
13. Two types of deployment
14. Bespoke Solutions “Ideas are worth nothing unless executed”, Derek Sivers c Daniel Moyo
15. Unused Models Many models never make it to production “Ideas are worth nothing unless executed”, Derek Sivers
16. Time to model deployment Model development = days to weeks Model deployment = weeks to never! “Ideas are worth nothing unless executed”, Derek Sivers
17. The technology
18. c Flaticon init.py task.py setup.py model.py Model Package
19. Repeatable, Reusable Process init.py task.py setup.py model.py Model Package
20. Data transformations Scikit-learn pipelines + custom transformers Transformation occurs in the model
21. Solution for other libraries too Add preprocess file to the package Image taken from Google Cloud documentation
22. Further customisation Custom scoring Custom prediction routines
24. Faster time to production c flaticon Fully Managed service
25. Not Quite!
26. Collaborative Project
27. ML Proxy (bespoke ML microservice)
28. Model Versioning
29. Monitoring - Model Performance
30. Monitoring - AI Platform Performance
31. Time to model deployment Model development = days to weeks Model deployment = hours to days “Ideas are worth nothing unless executed”, Derek Sivers
32. How to scale Use Cases and Buy in (Input Team Deployment
33. The right kind of independence c flaticon Data Scientists have full ownership over models
34. The right kind of independence c flaticon Data scientists work closely together
35. The right kind of independence c flaticon But they also work closely with other teams
36. Use cases and buy in c flaticon Focus on problems to solve
37. Use cases and buy in c flaticon Don’t start in the highest value area
38. Use cases and buy in Deploy a first version (not the best) as fast as possible
39. Test and learn Photo by Alex Kondratiev on Unsplash Use cases and buy in
40. Thank you @vickdata