Too many ideas, too little data

How to overcome the cold start problem when building data driven products and services

Transcript

1. Too many ideas, too little data How to overcome the cold start problem when building data driven products and services Thomas Pawlitzki & Markus Nutz ML Conference 2018
2. Why this talk? 2
4. 85% of all big data strategies fail according to a Gartner poll 4
5. We suggest … identify problem check if ML helps to solve it look, collect, measure data iterate solution and make it better … and we would like to show this at some examples ... 5
6. Who we are? 6
7. Thomas & Markus Dat e t Sur M t o l i t c li , co r @thopaw Sof r E in Sof r A hi t te -lo g, ba t l, ki -su n , t e k @markusnutz
8. What we do 8
9. The idea of an insurance https://en.wikipedia.org/wiki/Insurance “Insurance is a means of protection from financial loss. It is a form of risk management, primarily used to hedge against the risk of a contingent or uncertain loss.” many people pay a small amount to hedge against a risk. If the risk happens the financial costs are covered. idea was already practiced by sailors in 14th century 9
10. The idea of an insurance 10
11. idea of insurance from point of insurer risk management under writing claim proces sing Solve ncy II § VVG
12. Challenges for insurances remain the same 12
13. Our Vision 13
15. Machine Learning to the rescue?! 15
16. ML models assist to make jobs easier by helping dynamic & elastic pricing early loss detection on-time insurance demand forecasts claim prevention ... fraud detection claim processing underwriting asset management pre-sort emails of customers ... conversion optimisation predictive marketing satisfaction / churn real-time bidding new insurance products ... error prediction scaling performance monitoring Risk & Prices Organisation Marketing IT systems software testing development assistants ... 16
18. Bike or not?! 18
19. Show pictures of bikes of customers Idea: show bike pictures of customers - rate - comment - ... Hall of Fame But what about - inappropriate pictures - insufficient rights - personality rights 19
20. Do not reinvent the wheel Is image recognition an already solved problem? Is it easily accessible? Is it affordable? DO NOT BUILD IT AGAIN 20
21. Do not reinvent the wheel Is image recognition an already solved problem? Is it easily accessible? Is it affordable? DO NOT BUILD IT AGAIN 21
22. Carbon check 22
23. Classifying a bike as carbon or not Problem: - cannot insure carbon bikes yet, annoying if happens Solution: - build a classifier with an accuracy of at least 80% icons designed by Smashicon from Flaticon 23
25. Fraud 25
26. Helping fighting fraud with IOT data Problem: - using an IoT device to not only reduce theft, but also to fight fraud and be able to offer cheaper prices to honest customers Takeaway: - Additional data sources can be great features for your models - Key is to collect useful data icons designed by Smashicon from Flaticon 26
27. 27 “In the morning I went outside my home where I locked my bike, but it was gone. This was on the 16th of August. My contract is 564467133.”
28. Damage Prevention / Risk reduction 28
29. Location based safety features Problem: - assess accident risk and risk of theft - using open street map data, official statistics and small inquiries Approach: - combine strength of traditional statistics with Deep Learning icons designed by Smashicon from Flaticon 29
30. Vandenbulcke, Thomas and Panis (2013) 30
32. Overcome the cold start problem 32
33. Team & Organisation - sponsorship of (top) management - experimentation is encouraged - focus on business objective - no silos, sit with the (IT) team 33 Product Owner Dev Ops Specialist Marketing UX Designer Data Engineer Data Scientist Front & Backend Developers
34. Technical requirements 34 - stable - reliable - accessible https://hackernoon.com/the-ai-hierarchy-of-needs-18f111fcc007
35. Structuring your data problem - can the problem be solved via machine learning? - or is thorough data analysis sufficient? - do we have data? can data be acquired? - what is the priority? 35 - where do I want to go? - how will I get there?
36. Tools & Techniques 36 https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcTUHwNyKbQBP0PyruuZZZmNhg5UZKAEXvN3Awb5x2yVo4nnq5Ud
37. think simple in technical / model, but end 2 end do not build something that nobody needs measure data learn quickly and steer/pivot Idea Experiment Data build prototype experiment & measure learn MVP cycle after Steve Blank & Eric Ries 37
38. What is an MVP? build not increments build useable products/services monitor usage and performance of the product/service improve your product/service when it is necessary https://www.linkedin.com/pulse/mvp-bike-car-fred-voorhorst/ 38
39. Other way to see MVPs DON’Ts Too many features in the first version at ML products Too high accuracy of model and also too many features Changing the purpose of the app at ML products Changing the purpose of the model (e.g. variable to predict of the model) https://anoda.mobi/what-is-minimum-viable-product-and-how-to-build-it-right/ 39
40. Thank you! thomas.pawlitzki@freeyou.ag @thopaw markus.nutz@freeyou.ag @markusnutz http://freeyou.ag - Build something that helps people - Structure your data problem - embed Data Science / Machine Learning in an effective way in your organisation - Build end-2-end, before starting to perfectionate your model freeyou 40
41. Credits Photo by Cameron Venti on Unsplash Photo by Tiffany Nutt on Unsplash Photo by Mikkel Bech on Unsplash Photo by Patrick Hendry on Unsplash Photo by Callum Wale on Unsplash Grégory Vandenbulcke, Isabelle Thomas, Luc Int Panis, 2013, “Predicting cycling accident risk in Brussels: A spatial case–control approach”, Accident Analysis and Prevention 62 (2014) 341– 357 MVP - https://medium.com/hackerlife/how-to-define-your-minimum-viable-product-dc7e118baec1 41