Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Presentatie Picnic, from data to insights, Daniel Gebler

521 views

Published on

Presentatie Picnic

Published in: Data & Analytics
  • Be the first to comment

Presentatie Picnic, from data to insights, Daniel Gebler

  1. 1. From Data to Insights @daniel_gebler @picnic
  2. 2. Our Value Proposition Groceries + Mobile HomeOn-TimeLowest Price + + + =
  3. 3. 0
  4. 4. WHY
  5. 5. EXPENSIVE
  6. 6. WAITING
  7. 7. CUMBERSOME
  8. 8. 1.5%
  9. 9. Producers Consumers Traditional Supermarket WAREHOUSE
  10. 10. PEOPLE WHO ARE REALLY SERIOUS ABOUT SOFTWARE MAKE THEIR OWN HARDWARE Alan Kay
  11. 11. PEOPLE WHO ARE REALLY SERIOUS ABOUT ECOMMERCE MAKE THEIR OWN LOGISTICS Picnic
  12. 12. Picnic Producers Consumers WAREHOUSE
  13. 13. From 0 to 1
  14. 14. Day 1
  15. 15. From 1 to n
  16. 16. Technology is overestimated on the short term but underestimated in the long term. Roy Amara
  17. 17. Challenges
  18. 18. 30 Articles in 3 Minutes The Mobile Shopping Challenge
  19. 19. Our Shopping Time Journey 0 50 100 150 200 250 300 350 400 Aug-16 Sep-16 Oct-16 Nov-16 Dec-16 Jan-17 Feb-17 Mar-17 Apr-17 May-17 Sessiontime(inseconds) 2 1 3
  20. 20. The Mobile Shopping Solution - Bulk recommendations • Set of 4, 8 or 12 articles • Buy all with a single tap • 1-click shopping for half of your basket • Purchase confidence >90% • Covering repetitive & boring items
  21. 21. Challenges of bulk recommendations Precision challenge • 90% single item precision  28% precision for 12 items • 90% precision for 12 items  99% single item precision • Factor 10 better recommendations required! Item challenge • Seasonal availability • Seasonal variations • Event-based buying patterns
  22. 22. Formalization of the Challenge PR(next= |hist=(( ,t-4),( ,t-7),( ,t-9)))
  23. 23. Input Hidden Layers Output Monday (wk -2) Friday (wk -2) Wednesday (wk -1) Tomorrow Solution 1: Deep Recurrent Neural Network (LSTM) • Item likelihood to buy • Cat likelihood to buy • Next 7 days • Item/cat buying interval • Order history (articles, dates) • Normalized quantities • Days between orders x 1 x 2 x 0 x 3 x 0 x 1 x 2 x 2 x 0 x2 x 1 x 1 Item - Item relations Item - Day relations Itemset - Day relations y2 x1 x2 x3 y3 y1 z2 z3 z1
  24. 24. 50%
  25. 25. Solution 2: RFM-based order prediction Get last 10 orders Select top items (80% orders) Rank by [freq, qty] Display best items (min 4, max 12) … | | | Filter by seasonality … … … … … …
  26. 26. Result: Big and Deep data for optimal order prediction 65% 70% 75% 80% 85% 90% 0 20 40 60 80 100 120 140 Precision Number of orders Big data (lack of depth) Big & Deep data Deep data (lack of breadth)
  27. 27. +20x +10x +11x From Technology to Business – Consumer Insights +1
  28. 28. Each week 100s of products 1,000s of suggestions 10,000s of interactions The Co-Creation Challenge
  29. 29. Step 1: Create Visibility, Encourage Accountability, Celebrate Success
  30. 30. Step 2: ML-based classification of product suggestions Customer input (free text) Picnic Retail Platform (storage) Force.com (processing & analytics) Zendesk (Customer feedback) Picnic Retail Platform (status update) Azure ML (NeuralNet classification)
  31. 31. Result: Auto-classification 3 out of 4 suggestions Insufficient training data Max 91% accuracy
  32. 32. Data Science is the MVP for AI Products
  33. 33. The Distribution Challenge 20 min window 99% on-time 1% no show
  34. 34. Formalization of the Challenge Tdrop = Carea + C1 + Cambient|chilled·Nambient|chilled + Cfrozen·Nfrozen + Tdelta
  35. 35. The Interface Challenge Everything Everytime Everywhere Convenient
  36. 36. 1. Dream Big, Act small 2. Mission first, Data as support 3. Data Science first, AI second 4. Launch first, Scale second 5. Great products come from small teams Learnings from a disruptive Scale-up
  37. 37. Creating a mobile super service @picnic @daniel_gebler

×