ºÝºÝߣ

ºÝºÝߣShare a Scribd company logo

Teaser Lingerie Recommendations with Van de Velde - Data Innovation Summit 2018 - Brussels

•
•
Python Predictions
Python Predictions

Teaser and some slides for the presentation by Kevin Heyman, Data Scientist at Van de Velde and Nele Verbiest, Data Scientist at Python Predictions at Data Innovation Summit, June 27, 2018, Brussels. In this story, Kevin and Nele explain how they have build a recommendation engine to predict your next lingerie purchase

1 of 20
WE RECOMMEND YOUR NEXT LINGERIE PURCHASE KEVIN HEYMAN NELE VERBIEST JUNE 27th, Brussels di-summit.com
THE STORY
?the challenge of this project lies in offering valuable product recom- mendations in a digital context
THE PROCESS
RECOMMENDATIONS a critical component of many recommendation systems is a similarity matrix, indicating similarity between items (=products)
Teaser Lingerie Recommendations with Van de Velde - Data Innovation Summit 2018 - Brussels
PRIMADONNA SUMMER PRIMADONNA MEADOW Two bras are similar if they are bought by the same women. in this matrix, similarity can be defined as the overlap in the women who buy two products: large overlap = similar
Two bras are similar if they are bought by the same women. ANDRES SARDA - NEPTUNE MARIE JO L’AVENTURE - TOM in this matrix, similarity can be defined as the overlap in the women who buy two products: small overlap = unsimilar
using the similarity matrix to generate recommendations
we’ll use the similarity matrix to detect items most similar to each previous purchase, and assign points to each recommendation, so we can recommend the item with the highest total score
1 2 4 5 6 8 9 7 10 3 OFFLINE EVALUATION this approach showed a significant improvement over recommending the most popular product (graph is accurate but exact figures are confidential)
MAIL TEST INCREASE IN CTO CONTROL TEST also in a real-life mail test, the recommendations proved to increase click-through rate significantly when compared to a control group offering the most popular product (graph is accurate but exact figures are confidential)
RECOMMENDATIONS CURRENT & OLD SEASON the previous approach only works for previous seasons, for which we have purchase data, but what if we don’t?
PRIMADONNA SUMMER PRIMADONNA MEADOW Two bras are similar if they are bought by the same women. ? MARIE JO ERIKA AW 2018 for a new season we don’t know who will purchase so we cannot calculate the overlap
PRIMADONNA DIVINE to solve this, we created a deep neural network to detect similarity between purchased items and the new collection we first trained this network to classify images from the new collection correctly
MARIE JO SAKURA next, we apply the network to previous purchases, so we use the model to predict similar items
MARIE JO - FLEUR MARIE JO L’AVENTURE - MAI example of a detected similarity: old season -> new season
PRIMADONNA TWIST – TOUGH GIRL PRIMADONNA TWIST - CABARET example of a detected similarity: old season -> new season
RECOMMENDATIONS NEW SEASON RECOMMENDATIONS CURRENT & OLD SEASON as a result, we are able to offer recommendations for existing seasons but also new seasons
WE RECOMMEND YOUR NEXT LINGERIE PURCHASE BUT WE DO MORE THAN THAT…

More Related Content

Teaser Lingerie Recommendations with Van de Velde - Data Innovation Summit 2018 - Brussels

  • 1. WE RECOMMEND YOUR NEXT LINGERIE PURCHASE KEVIN HEYMAN NELE VERBIEST JUNE 27th, Brussels di-summit.com
  • 3. ?the challenge of this project lies in offering valuable product recom- mendations in a digital context
  • 5. RECOMMENDATIONS a critical component of many recommendation systems is a similarity matrix, indicating similarity between items (=products)
  • 7. PRIMADONNA SUMMER PRIMADONNA MEADOW Two bras are similar if they are bought by the same women. in this matrix, similarity can be defined as the overlap in the women who buy two products: large overlap = similar
  • 8. Two bras are similar if they are bought by the same women. ANDRES SARDA - NEPTUNE MARIE JO L’AVENTURE - TOM in this matrix, similarity can be defined as the overlap in the women who buy two products: small overlap = unsimilar
  • 9. using the similarity matrix to generate recommendations
  • 10. we’ll use the similarity matrix to detect items most similar to each previous purchase, and assign points to each recommendation, so we can recommend the item with the highest total score
  • 11. 1 2 4 5 6 8 9 7 10 3 OFFLINE EVALUATION this approach showed a significant improvement over recommending the most popular product (graph is accurate but exact figures are confidential)
  • 12. MAIL TEST INCREASE IN CTO CONTROL TEST also in a real-life mail test, the recommendations proved to increase click-through rate significantly when compared to a control group offering the most popular product (graph is accurate but exact figures are confidential)
  • 13. RECOMMENDATIONS CURRENT & OLD SEASON the previous approach only works for previous seasons, for which we have purchase data, but what if we don’t?
  • 14. PRIMADONNA SUMMER PRIMADONNA MEADOW Two bras are similar if they are bought by the same women. ? MARIE JO ERIKA AW 2018 for a new season we don’t know who will purchase so we cannot calculate the overlap
  • 15. PRIMADONNA DIVINE to solve this, we created a deep neural network to detect similarity between purchased items and the new collection we first trained this network to classify images from the new collection correctly
  • 16. MARIE JO SAKURA next, we apply the network to previous purchases, so we use the model to predict similar items
  • 17. MARIE JO - FLEUR MARIE JO L’AVENTURE - MAI example of a detected similarity: old season -> new season
  • 18. PRIMADONNA TWIST – TOUGH GIRL PRIMADONNA TWIST - CABARET example of a detected similarity: old season -> new season
  • 19. RECOMMENDATIONS NEW SEASON RECOMMENDATIONS CURRENT & OLD SEASON as a result, we are able to offer recommendations for existing seasons but also new seasons
  • 20. WE RECOMMEND YOUR NEXT LINGERIE PURCHASE BUT WE DO MORE THAN THAT…