Building an Incrementally
Trained Global Deep Learning
Recommender System Model
Anoop Deoras, Ko-Jen (Mark) Hsiao
adeoras@netflix.com
MLConf, San Francisco
11/08/2019
@adeoras
~150M Members, 190 Countries
● Recommendation Systems are means to an end.
● Our primary goal:
○ Maximize Netflix member’s enjoyment of the selected show
■ Enjoyment integrated over time
○ Minimize the time it takes to find them
■ Interaction cost integrated over time
Personalization
● Personalization
Everything is a recommendation!
Ordering of the titles in each row is personalized
Selection and placement of the row types is personalized
Profile 1
Profile 2
Personalized
Images
Personalized
Messages
IMPRACTICAL TO
SHOW EVERYTHING
We Personalize our recommendation!
This Talk Answers: HOW ?
Basic Intuition behind Collaborative Filtering
● Imagine you walked into a room full of movie enthusiasts, from all over
the world, from all walks of life, and your goal was to come out with a
great movie recommendation.
● Would you obtain popular vote ? Would that satisfy you ?
Basic Intuition behind Soft Clustering Models
● Now consider forming groups of people with similar taste based on the
videos that they previously enjoyed.
Basic Intuition behind Soft Clustering Models
● Describe yourself using what you have watched.
● Try to associate yourself with these groups and obtain a weighted
“personalized popularity vote”.
Distribution over the topics and over the videos
0.01
0.63
0.22
0.15
Topic Models (Latent Dirichlet Alloc)
K
U
P
α θ φt v
β
Total
Topics
Taste
Convex Combinations of
topics proportions and movie
proportions within topic
OUR ALGORITHMS ARE GLOBAL AND THEY HELP LOCAL
STORIES BE HEARD GLOBALLY
GLOBAL ALGORITHMS
foster
GLOBAL COMMUNITIES
Thanks to Sudeep Das for contributing this beautiful slide.
Country Context in LDA models
Users in Country A play both Friends and HIMYM Users in Country B cannot play both Friends and
HIMYM
Country A catalog Country B catalog
Model is forced to split HIMYM plays.
topic k
Outcome: Parameters are being consumed to explain catalog differences.
topic j
Topic with
high mass
on Friends
and HIMYM
Topic with
high mass
on HIMYM
Thanks to Ehtsham Elahi for contributing this slide.
Catalogue Censoring in Topic Models
K
U
P
α θ φt v
β
Total
Topics
Taste
c
Censoring
pattern
m
Global Recommendation System for Overlapping Media Catalogue, Todd et.al., US Patent App
ALGORITHMS NEED TO CAPTURE THE TREND
Time context in Topic Models
K
U
P
α θ φt v
β
Total
Topics
Taste
m
Observed
time
µ
Topics over Time: A Non Markov Continuous-Time Model fo Topic Trends. , Wang et.al., KDD 2006
Fully contextualizing Topic Models
K
U
P
α θ φt v
β
Total
Topics
Taste
m
Observed
time
µ
c
Censoring
pattern
m
SIMPLE !
IMPRACTICAL TO
SCALE
Differentiation
manual
Time Consuming
Poor Scaling
Symbolic
Time Efficient
Excellent for
scaling
Gift of Deep Learning: Automatic Differentiation
Variational Autoencoders
zu
u
Taste
fθ
𝞵 𝞼
u
Encoder
Decoder
fѰ
fѰ
DNN
Soft-max over entire vocabulary
Variational Autoencoders for Collaborative Filtering, Liang et al. WWW (2018)
ReLU
ReLU
User
Representation
Feed
Forward
Country Catalogue
Country
● Create a censored
mask with out of
catalogue videos
● Mask the output layer
(logits)
● Use the masked layer
for cross entropy loss.
How to do country
contextual modeling ?
ReLU
ReLU
User
Representation Country
Feed
Forward
Country Catalogue
Save Model Energy in
Learning Catalogue
Differences
ReLU
ReLU
User
Representation
Country
Feed
Forward
Adding Time is Easy too
Time at
Serving
Time to train is large ! Catalogue changes quite fast
DL Model
P( | U, C, T)
Cannot estimate as
not in our model’s vocabulary
Incrementally Train the Models
time
Fully Trained Model Additional Nodes and
Parameters
RECIPE
1. CENSOR
2. ADD CONTEXT VARIABLES TO THE MODEL
3. DO; EVERY FEW DAYS
a. TRAIN A WARM START MODEL WITH (1 & 2)
4. DO; EVERY FEW HOURS
a. TAKE THE MODEL FROM (3)
b. ADD NEW EMBEDDINGS
c. ADD NEW PARAMETERS
d. FINE TUNE
THANK YOU !
Questions ?
Anoop Deoras, Ko-Jen (Mark) Hsiao
adeoras@netflix.com
@adeoras
Sincere thanks to a lot of my Netflix Colleagues: Aish Fenton, Dawen Liang and
Ehstham Elahi for contributing to the ideas discussed here.

Anoop Deoras - Building an Incrementally Trained, Local Taste Aware, Global Deep Learned Recommender System Model