polyvore outfit recommender system
DESCRIPTION
Presentation for ISR project about Polyvore recommender systemTRANSCRIPT
OUTFIT RECOMMENDATIONA collaborative fashion recommendation system for e-commerce
Maria NadejdeLamar Payson
Information Search and RetrievalSpring 2010FUB
Characteristics of the Fashion Domain Large feature set Diversity of consumer taste (styles) Importance of user demographics Short product life (trends, campaigns)
Polyvore - A social website for Fashion Clothing items
Individual clothing items imported from the web by the users
Price, and link to original site automatically recorded
Item tags are generated by the user
Polyvore – Outfits and Collections•Users create outfits (called “Sets”) from clothing items that have been imported
•Users may add outfits into collections of outfits
•Items, Outfits, and Collections may all be “liked” by the user
•So…• Outfits: sets of
items• Collections: sets
of sets of items
Note: Clothing items and outfits have tags, collections do not
Polyvore - Groups•Users may join groups of like minded users with similar tastes
•Note: we may view group membership as similar to demographic similarity.
Outfit Recommendation
GOAL: Recommend new outfits that a given user will like
Collaborative: From the set of all users U, find users
similar to me based upon (one or more of): Liking clothing items that I like Liking outfits that I like Liking collections that I like Being in the groups that I am a member of
But…Sparsity
Users: 1.2 million Clothing items: >500,000 (a
conservative estimate)
Outfits: 30,00 outfits created daily
The “Top” outfit for April 29th had only 385 “likes”
Outfit Recommendation (cont.) Content Based:
Tag set
Outfits liked by
user
Clothing items
have
have
Similar Outfits
members of
CollectionsCollections
liked by user
have
Outfits
Similar Clothing
items
members of
members of Clothing
items liked by user
Have
Members
of
Members
of
Outfit Recommendation (cont.) Hypothesis: The large number of
different styles creates highly segmented consumer interests. Collaborative filtering is most desirable but
may be precluded by sparsity of data Idea 1: Use the hierarchical structure of data
on Polyvore (and collaborative-via-content?) to overcome data sparsity
Idea 2: If needed, switch to a cascade method using content based filtering and demographic information (group membership?)