Transcript RPEC2002

Recommender Systems and
Product Semantics
Rayid Ghani & Andy Fano
Accenture Technology Labs
Workshop on Recommendation & Personalization in E-Commerce
May 28, 2002
Who we are?
Accenture Technology Labs

R&D Group for Accenture

~ 40 researchers in Chicago, Palo Alto
(California) and Sophia Antipolis (France)

Research in Data Mining, Machine Learning,
Ubiquitous Computing, Wearable Computing,
Language Technologies, Virtual & Augmented
Reality, Collaborative Workspaces…
What Does a Transaction
Mean?
Terabytes of transaction data.
But what does any one transaction mean?
What does it tell us about the customer?
Example: Apparel
Transactional information captured by retailers:
 Date of Purchase
 SKU
 Price
 Size
 Brand
But what does this tell me about the customer who
bought it?
Product Semantics:
What does a product mean?
What does this shirt say about
her?
Is it conservative or flashy?
Trendy or classic?
Formal or casual?
Where would we get this
information?
Where do people get this information?
Marketing
Product Companies and
Retailers spend fortunes
telling customers what
their products mean.
Our idea:
Build a system that
analyzes marketing
texts to infer these
attributes.
Example
From the Macy’s web site:
DKNY Jeans Ruched
Side-Tie Tee
Get back to basics with a fresh
new look this season. The
Ruched Side-Tie Tee has a
drawstring tie at left hip with
shirred detail down the side.
Stretch provides a flattering,
shapely fit. V-neck.
Training the System
Product
Descriptions
Domain
Experts
Product descriptions
marked up with
attribute values
Supervised
Learning
Algorithm
Learned
Statistical
Models
Inferring Attributes via Text
Classification


Build one classifier per attribute type
Simple statistical classifier – Naïve Bayes
Multinomial model (McCallum & Nigam 1998)


For all words (description) and attribute values:
 calculate P(word | attribute value) using the
manually rated items
Given a new item description:
 Calculate P(attribute value | item description) for all
attribute values
 Use Maximum Likelihood
Semi-supervised Learning



Lot of product descriptions available for
minimal cost
Labeling them is expensive
Apply magical algorithms that combine labeled
and unlabeled data for classification

EM (Nigam et al. 1999), Co-Training (Blum &
Mitchell 1999), Co-EM (Nigam & Ghani), ECo-Train
(Ghani, 2002)
The EM Algorithm
Estimate
labels
Learn from
labeled data
Naïve
Bayes
Probabilistically add
to labeled data
A Peek at the Learned Models
Not Conservative
(Flashy)
Extremely
Conservative
rose
special
leopard
chemise
straps
flirty
spray
silk
platform
lauren
ralph
breasted
seasonless
trouser
jones
sport
classic
blazer
Bias Slip Dress
The perfect black dress gets flirty and feminine in the
bias-cut slip dress with sheer ruffled cap sleeves. A
low, scoop neck and back is ultra-flattering while a
draped, romantic fit reveals total elegance.
Lauren Single-Breasted Blazer
Sporty elegance and classic Gatsby-esque styling are
captured in this impeccably designed single-breasted,
three-button blazer from Lauren by Ralph Lauren.
With traditional notch collar, signature button
hardware, front flap pockets, and signature crest on
left breast pocket.
A Peek at the Learned Models
Informal
formal
jean
tommy
denim
sweater
pocket
neck
tee
hilfiger
jacket
fully
button
skirt
lines
seam
crepe
leather
Polo Jeans Co. Muscle Logo Tee
Strut your stuff in the Muscle Logo Tee. Flattering on
the arms with a close-to-the-body fit, classic crewneck
and shimmery logo print with stars. A sporty new basic
for your tee collection.
BLACK TRIACETATE JACKET
A fresh alternative to classic suiting. Wear
open for cardigan effect, buttoned for a
clean look. Hidden placket with four tonal
buttons and a hook-and-eye closure at
the collar. Falls to hip. Lined.
A Peek at the Learned Models
Loungewear
Partywear
chemise
silk
kimono
calvin
klein
august
lounge
hilfiger
robe
gown
rock
dress
sateen
length:
skirt
shirtdress
open
platform
plaid
flower
ABS by Allen Schwartz Asymmetrical Dress
Just for the party girl with a big feminine streak.
A ruffled one-shoulder cuts diagonally across the
front and back. Accented with a rhinestone detail
on the shoulder.
A Peek at the Learned Models
Juniors
Extremely Sporty
jrs
dkny
jeans
tee
collegiate
logo
tommy
polo
short
sneaker
sneaker
camp
base
rubber
sole
white
miraclesuit
athletic
nylon
Mesh
DKNY Jeans Jrs. Mesh Jersey Sweater
An innovative take on the football jersey, the
see-through mesh sweater is a fashion
favorite among the sporty set. Denim appliqué
Populating the Knowledge Base
New
Product
Descriptions
Learned
Statistical
Models
Product descriptions
automatically marked up
with attribute values
Product Semantics
Knowledge Base
Recommender System
Retailer’s
Web Site
Extracted
Descriptions
of Products
Browsed
Product Semantics
Knowledge Base
Learned
Statistical
Models
Evolving
User Profile
Advantages over Traditional
Recommendation Systems
This approach provides us some of the underlying
attributes that characterize a customer’s preference.
We can therefore begin to explain the preference rather
than simply rely on the co-occurrence of purchases (e.g.
people who bought x also bought y).
This helps with:
 Handling new products/rapidly changing products
 Low Frequency Products
 Cross Category Recommendations
Cross-Category
Recommendations




Difficult for collaborative filtering and contentbased systems
Build a model of the user - personality,
stylistic attributes
Taste in clothing might also be suggestive of
taste in other products, say furniture and
home decoration
Create models for different product classes
and create mappings among these models
Summary



“Understand” a product and hence the
customer
Use Text Learning (supervised and semisupervised) to abstract from product
(description) to subjective, domain-specific
features
Effective for new (and low frequency)
products and for cross-category
recommendations