Data and Personalization: The Marketplace Holy Grail
•
4 j'aime•789 vues
Personalization is the key to the future of marketing in e-commerce. Extracting meaningful information from the data sources is the first step in building learning models that are used to craft a personalized experience in Etsy. Etsy is an online market place for artisans selling unique handcrafted goods, and vintage wares. Kamelia is a data scientist in Etsy for the past two years. In this talk, she’ll discuss data extraction and machine learning techniques behind user, shop and listing recommendations that create a personalized experience for Etsy users!
Recommandé
Recommandé
Contenu connexe
En vedette
En vedette (8)
Plus de Extract Data Conference
Plus de Extract Data Conference (18)
Dernier
Dernier (20)
Data and Personalization: The Marketplace Holy Grail
- 2. Data and Personalization: The Marketplace Holy Grail ! Kamelia Aryafar (@KAryafar) , Data Scientist @ Etsy ! https://github.com/karyafar/Extract2015
- 3. Data Science 3 Kamelia Aryafar Diane Hu Andrew Clegg Rob Hall Josh Attenberg Corey Lynch Caitlin Cellier
- 4. Etsy is an online marketplace where people connect to buy and sell unique goods: Handmade, vintage, or craft supplies Etsy Overview 4 ColoredStars Rome , Italy
- 5. Etsy Overview 5 ~21M Active Buyers 1.4M Active Sellers ~35M Listings ! ! Stats as of March 31, 2015 People buy and sell on Etsy from nearly every country in the world.
- 6. In a 2014 survey of Etsy buyers, 92% agreed that Etsy offers products they cannot find elsewhere. 6
- 7. Unique Items, Unique Challenges
- 10. Etsy Overview 10
- 11. ! ! ! Favorite an item or shop, and add to collections with coherent theme/style ! Follow another user with similar style/interest ! ! ! ! Browse (Unintentional) ! Search (Intentional) Etsy Overview How do people typically use Etsy? 11 ! ! ! Purchase an item ! ! ! ! ! Marketing Emails Including Recommendations ! ! ! ! ! Pinterest Facebook Search Engines ! … 1 2
- 12. Data, Data, Data
- 13. Recommendations on Etsy 13 ITEM RECS USER REC SHOP REC
- 18. Empty C art Recommendations 18
- 19. Personalization
- 20. • Collect Data • Scrub • Model • Generate Recommendations Personalization 20
- 21. Matrix Factorization ( MF ) 21Slide Credit: Rob Hall
- 22. Matrix Factorization ( MF ) 22Slide Credit: Rob Hall
- 23. Matrix Factorization ( MF ) 23Slide Credit: Rob Hall
- 24. L ocality S ensitive Hashing ( LSH ) 24Slide Credit: Rob Hall
- 25. Latent Dirichlet Allocation ( LDA ) 25 Discover popular styles on Etsy as a distribution over items Represent each user as a distribution over popular styles, i.e. “style profile” ... V K “geometric” “mid-century” “surreal” . . . K 0.38 0.13 0.02 ... “geometric” “mid-century” “surreal”= Slide Credit: Diane Hu
- 26. Discovering User Styles 26 A N I M A L S G E O M E T R I C M I D - C E N T U RY M O D E R N T E N TA C L E S Different styles discovered by LDA Slide Credit: Diane Hu
- 27. Generating Recommendations 27 S T Y L E # 4 2 8 S T Y L E # 6 5 5 S T Y L E # 5 4 S T Y L E # 8 7 U S E R ’ S FAV O R I T E S I T E M R E C O M M E N D AT I O N S Slide Credit: Diane Hu
- 28. What We Have Learned
- 29. Impact 29 A/B Testing ! •Variant: LDA or MF personalized recommendations •Control : No recommendations ! • Significant increase in all business metrics !
- 30. Personalized content best engages users and enhances business metrics. 30
- 31. More Data, More accurate models, Better recommendations 31