4. In this talk
• Fashion Domain Characteristics
• Advice & Recommendation Reasons
5. Fashion Domain
Items: “Fast Fashion”
• Short lifetime of items: a couple of months on average
• Large retailers release new garments daily
6. Fashion Domain
Users
• Taste changes over time
• Seasons complicate things
• Trends can drastically change user preferences very quickly!
7. Fashion Domain
What this means for recommenders
• Item catalogue changes rapidly
• High sparsity
• Use content data
• We need recommenders that can recognise changes in user
preference and respond quickly
8. Advice & Recommendation Reasons
Why do we need an advice component
• Users are looking for guidance and validation that
something will or will not work for them
• Users might not know what fits them best
• Fashion confidence is a big factor – out of comfort zone
recommendations
9. Advice & Recommendation Reasons
Challenges
• Can’t blindly trust historical interaction data
• Need additional information on users and garments
• Need expert knowledge
• Need to communicate in a user-understandable way
10. Advice & Recommendation Reasons
How we do this
• Users fill out questionnaire
• We label garments
• Stylists encode fashion rules
13. “Your key shapes:
The key to dressing your slim silhouette is to disguise your broad
shoulders. Clothes that add flattering volume to your lower half
help to balance out your proportions.
Look for lower necklines that soften you shoulder line. Avoid fussy
shoulder details that add bulk to this area.”
14.
15. Advice & Recommendation Reasons
• Recommendation reasons have to make sense to the user
• The point of reasons is not to justify “why are you showing
me this”, it is to answer “why is this item good for me”
• Give useful information to the user
16. Summary
• Item catalogue and user preferences change over time,
faster than in other domains
• It’s not just about product recommendations, users want
advice
• Make recommendation reasons useful to the user
• User-Item interaction data alone is not enough
17. Learnings
• Start from domain rather than algorithm
• Human-understandable features are great!
• allow for applications that are useful to the user, i.e. advice and
recommendation reasons
• allow for more involvement of domain experts and more rapid
iteration on algorithm development
18. Thanks!
• We are hiring!
• Open for collaboration with academia!
• nick@dressipi.com