Visual Discovery Daniel Hill Amazon Personalization Sciences Palo Alto CA wwwamazoncom stream RecSys 2016 3 Beyond Accuracy Challenges for visual browsing Freshness Surface new products without degrading stream quality ID: 525145
Download Presentation The PPT/PDF document "Adaptive, Personalized Diversity for" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Adaptive, Personalized Diversity for
Visual Discovery
Daniel Hill
Amazon Personalization Sciences
Palo Alto,
CASlide2
www.amazon.com
/streamSlide3
RecSys 2016
3Beyond Accuracy: Challenges for visual browsingFreshnessSurface new products without degrading stream quality
Diversity
Create
visual interest
Hedge
against
uncertainty
in customer
intent
Seasonality
Adapt quickly to trends in fashion
Personalization
Increase relevance to customer’s styleSlide4
RecSys 2016
4Beyond Accuracy: Challenges for visual browsingFreshnessSurface new products without degrading stream quality
Diversity
Create
visual interest
Hedge
against
uncertainty
in customer
intent
Seasonality
Adapt quickly to trends in fashion
Personalization
Increase relevance to customer’s styleSlide5
RecSys 2016
5Beyond Accuracy: Challenges for visual browsingFreshnessSurface new products without degrading stream quality
Diversity
Create
visual interest
Hedge
against
uncertainty
in customer
intent
Seasonality
Adapt quickly to trends in fashion
Personalization
Increase relevance to customer’s styleSlide6
RecSys 2016
6Beyond Accuracy: Challenges for visual browsingFreshnessSurface new products without degrading stream quality
Diversity
Create
visual interest
Hedge
against
uncertainty
in customer
intent
Seasonality
Adapt quickly to trends in fashion
Personalization
Increase relevance to customer’s styleSlide7
RecSys 2016
7Explore-exploit in product scoring
Product ID
Brand
Price
Category
100352463
Jacket
Calvin Klein
$98
Bayesian Linear
Probit
Regression
Click
or
No-click
Product attributes,
x
Attribute weight
distributions, w
Product Score =
F
(
Σ
i
w
i
x
i
)
Thompson Sampling
C
ustomer
actions
W
i
~ N(0,1)Slide8
RecSys 2016
8Ranking by product s
core alone Slide9
RecSys 2016
9Diversification applied to product categoriesCategory: (
Department
,
Product Type
,
Price Range
)
womens
-
dress
-
high
Submodular utility
Stream
utilitySlide10
RecSys 2016
10Submodular utility function using adaptive category weights
utility(
D
) =
Σ
k
w
k
log[1 +
c(
k
,
D
)
] +
s(
k,D
)
Number of products in category k
Weight for category k
Total score
of products in category kSlide11
RecSys 2016
11Submodular utility function using adaptive category weights
utility(
D
) =
Σ
k
w
k
log[1 +
c(
k
,
D
)
] +
s(
k,D
)
Number of products in category k
Weight for category k
Total score
of products in category k
Category weights are smoothed
click-through-rates
over a
rolling-window
w
k
= (
clicks
k
+ α
k
) / (
views
k
+ α
k
+ β
k
) Slide12
RecSys 2016
12Personalization of category weights from customer actions
Customer
clicks
c
u
~
Multinomial(
w
u
)
weights,
w
u
~
Dirichlet
(
w
)
Personalized weights,
E[
w
u
]
= (
c
u
+
w
)
/ |
c
u
+
w
|1Slide13
RecSys 2016
13
×
Category-category correlations
Customer preference vector
Smoothed
customer preference vector
Diffusion of customer preferences via category correlationsSlide14
Live experimental results
ComponentLift in product clicks
Submodular
vs. proportional diversity
+2.2
%
Adaptive
category weights
+10.4
%
Personalized
category weights
+6.1
%
RecSys
2016
14Slide15
RecSys 2016
15Thank you!
Choon
Hui
Teo
Houssam
Nassif
Mitchell Goodman
Vijai
Mohan
S. V. N.
Vishwanathan
Sriram
Srinavasan
Daniel Hill
Amazon Personalization
Sciences
Palo
Alto,
CASlide16
RecSys 2016
16Slide17
Stream ranker workflow
RecSys 2016
17
Event Logs
Click Model
Updater
Seasonal
Weights
Personalized
Weights
Category
Diffusion Matrix
Catalog
Ranked stream
Diversified stream
Rediversifier
BrowserSlide18
RecSys 2016
18Category-category CorrelationsDiffuse preference for category
A
to category
B based on the
correlation
between A and B
“clicked A also clicked B”