Realtime Contextual
Product
Recommendations…
… that scale and generate revenue
Pallav Agrawal
Director, Data Science
Levi Strauss & Co.
I WANT TO KNOW YOU
“38% of consumers
say they won’t
return to an online
retailer that
recommends things
that don’t make
sense for them.”
— Bazaarvoice Consumers
Survey, 2018
“Tanner was super friendly & helpful from the
moment we walked in the door! He was very
knowledgeable about the different styles and helped
me find exactly what I was looking for, even though I
wasn’t expecting to find it. He made the experience
so much better and I found some amazing pieces!
Thank you Tanner!”
Qualities of a Good
Recommender System
Makes the customer journey intuitive and
frictionless
Best reflects and communicates your
organization’s brand values
Focuses on what consumers want, not on
who they are
Uses both implicit behavioral markers and
explicit input
Provides relevant information in a timely
manner to expedite decision making
Once Upon a Time...
Why? Based on what
information?
Expensive and discounted
products shown next to each other
random female jean appears
Initial Architecture
24h
3rd
Party
API
Architecture v2
Architecture v3
Architecture v4
Popularity-Based
Recommendation
Model
source
Clearly articulated
source of information
Location based targeting determines
bestselling products in a colder climate
Association Rules
Mining based
Recommendation
Model
source
Clearly articulated
source of information
Mix of female tees and jeans to
recommend outfitting options
Item to Item
Similarity Based
Recommendation
Model
source
Clearly articulated
source of information
Person browsed girls tees with a prominent Levi’s
logo on a monochrome background
Dark shade tee to create a
breakpoint and add visual interest
Product Attributes:
Name: Wedgie Fit Straight Taper Jeans
Gender: Women
Product Type: Jeans
Material: 100% Cotton
Color: Light Wash
Rise: High Rise
Size Group: Women’s Regular
Fly: Button Fly
Fit: Straight, Wedgie
Merchant Badge: Waterless
Stretch: Non-Stretch
Distress: Distressed
Price: $110
Leg Opening: Tapered
Ankle: Cropped
Obtaining Image Embeddings
Obtaining Image Embeddings
Sparse Feature Matrix
Curse of Dimensionality
Obtaining Image Embeddings
Fashion Image
Embeddings
[0.01359, 0.00075997, ..., 1.0048, 0.06259]
[-0.24776, -0.12359, 0.20986, ..., 0.079717]
[-0.35609, 0.21854, 0.080944, ..., -0.35413]
Transfer Learning with Custom Dataset
BackpropFrozen Weights
Document Similarity using Text Embeddings
Document Similarity using Text Embeddings
High Rise Highrise
Whiskers Whiskers
Custom Word Embeddings
source
Collaborative
Filtering-Based
Recommendation
Model
source
Collaborative Filtering
Clearly articulated
source of information
Recommended Tees due to past
purchase of Jeans
Women’s tee shown because
of previous mix-gender
product purchases
“You have to start with the customer experience and work
backwards to the technology”
Steve Jobs
Realtime Contextual Product Recommendations…that scale and generate revenue - Pallav Agrawal

Realtime Contextual Product Recommendations…that scale and generate revenue - Pallav Agrawal

Editor's Notes

  • #4 How many of these terms sound familiar to you
  • #5 40% of spotify users use it, Streamed 10s of billion tracks
  • #6 35% of amazon.com revenue comes from it’s recommendations
  • #7 Basically everywhere, people expect good recos
  • #8 Levi’s store stylists are trained to be really good at helping people who are looking for recos. How do they do that.
  • #9 Here;s what makes for a good reco. HMW enable this online?
  • #12 Recommendations are effectively a problem of learning how to rank, so we moved to an architecture that would treat recommendations like a search problem. So we turned a trusted search engine backend: ES But the problems were far too many, constantly reindexing the same documents is bad for ES, and recommendations weren’t very relevant. Also, we wanted more flexibility in terms of what products get shown.
  • #26 Explaining similarity, NLP embeddings and TF-IDF, along with relevant attribute selection from merch, what’s the data source
  • #29 Resnet + added layers, resnet frozen, backprop on new layers
  • #30 tf-idf
  • #31 tf-idf