style conditioned recommendations - recsys€¦ · 3 © overstock diversification in recs users may...
TRANSCRIPT
StyleConditioned Recommendations
Murium Iqbal,
Timothy Anderton, Kamelia Aryafar
2
© Overstock
Filter Bubble
Current History
Horror
Thriller
Obvious Recommendations
Horror
Thriller
Mystery
What about?
Rom-com
Sci-fi
Fantasy
Action
3
© Overstock
Diversification In Recs
Users may want to explore outside of previous history
–New styles of clothing
–New genres of media
–New taxonomies of items
–New styles of furniture
4
© Overstock
Diversification In Recs
5
© Overstock
Ways to gather new style interests
• Self identification
• Quizzes for the user
• Stylist feedback
• Learning over favorites
6
© Overstock
Diversification In Recs
Can't filter recs by style
Attribute stuffing in labels
– Modern
– Coastal
– Glam
– Bohemian
Can't use top selling by style
Lacks personalization
7
© Overstock
Diversification
• Create a model that can diversify in a personalized way
• Can learn user styles from limited data
• Can make use of explicit feedback from user or expert feedback
8
© Overstock
Our Approach
• Use Conditional Variational Autoencoders to model user's current behavior
• Use another encoder to learn user's style profile
• Incorporate feedback to inject new styles into the user's recommendations
9
© Overstock
Conditional Variational AutoEncoder
Encoder Decoder
µ
Σ x
+
α ~N(0,1)
User
Click
Matrix
XC'
zC
User
Style
Profiles
XC
zT
10
© Overstock
Conditional Variational AutoEncoder
Encoder Decoder
µ
Σ x
+
α ~N(0,1)
User
Click
Matrix
XC'
zC
User
Style
Profiles
XC
zT
New
Style
Profiles
11
© Overstock
Style Injection
•Concat style labels to inputs
•Enable style transfer techniques
12
© Overstock
Style Injection
Encoder Decoder
µ
Σ x
+
α ~N(0,1)
User
Click
Matrix
XC'
zC
User
Style
Profiles
XC
zT
New
Style
Profiles
[0,0,0,0,0,0,1,0]
Modern
[0,0,0,0,0,1,0,0]
Bohemian
13
© Overstock
Style Injection
•Transfer user's preferences from one style to another
•Need to learn user's current profile
•Requires user style profiles to be interpretable
14
© Overstock
Learning User Style Profiles
• Limited labeled items (8k out of 41k)
• Need to extend item labels to users
15
© Overstock
Learning User Style Profiles
• Create a training dataset from labeled style
items
• Features should represent user's history
• Labels should represent distribution of style
within history
16
© Overstock
Learning User Style Profiles
• Sample sets of items from user click data
• Limit to items with style labels
• Extend item labels to users by averaging and thresholding to
create multivariate bernoulli label
• This allows us to maintain associations b/w styles which exist
in user click data
17
© Overstock
Learning User Style Profiles
Each item can be seen as the average of
word2vec embeddings of their title and
attributes
Each user can be thought of as the average of representations of all the items
they have interacted with
Users with different number of clicks will
have different distributions of embeddings
Limit each user to 5 items
18
© Overstock
Learning User Style Profiles
Use this as input to an MLP to create a
user profile
Each dimension of the profile corresponds to a predefined style
Profiles are directly interpretable, and can
be updated at decoder
19
© Overstock
Style Conditioned Recs
20
© Overstock
Style Conditioned Recommendations
•Can be taken from explicit feedback
•Can be taken from most popular item's style profile
•In absence of new style profile can be taken from last 5 items
21
© Overstock
Style Injection on Recommendations
Validating by injecting different styles into
a user's recommendations. Each injection
is taken as a single one-hot encoded style
Once style is injected learn the profile on the injected recommendations
The presence of the injected style
increases in the recommendations