For online businesses, recommender systems are paramount. There is an increasing need to take into account all the user information to tailor the best product offer, tailored to each new user. Part of that information is the content that the user actually sees: the visuals of the products. When it comes to products like luxury hotels, pictures of the room, the building or even the nearby beach can significantly impact users’ decision. In this talk, we will describe how we improved an online vacation retailer recommender system by using the information in images. We’ll explain how to leverage open data and pre-trained deep learning models to derive information on user taste. We will use a transfer learning approach that enables companies to use state of the art machine learning methods without needing deep learning expertise.

1. Extracting information from images using
deep learning and transfer learning.
Application to personalized recommendations.

2. Introduction and Context
Iterative building of a
recommender system
Labeling Images
Pragmatic deep learning for
dummies
Post Processing
AKA: Image for BI on steroids
Outline
Results
More images !

3. Dataiku
• Founded in 2013
• 90 + employees, 100 + clients
• Paris, New-York, London, San Francisco,
Singapore
Data Science Software Editor of Dataiku DSS
DESIGN
Load and prepare
your data
PREPARE
Build your
models
MODEL
Visualize and share
your work
ANALYSE
Re-execute your
workflow at
ease
AUTOMATE
Follow your production
environment
MONITOR
Get predictions
in real time
SCORE
PRODUCTIO
N

4. Client Key Figures
E-business vacation retailer
Negotiate the best prize for their clients
Discount luxury
Sale Image is paramount
Purchase is impulsive
18 Millions of clients.
Hundreds of sales opened everyday

5. Specificities
Highly temporary sales
-> Classical recommender system
fail
-> Time event linked (Christmas, ski,
summer)
Expensive Product
-> Few recurrent buyers
-> Appearance counts a lot

6. Iterative Building of a Recommender System

7. Basic Recommendation Engines

8. Other Factors

9. One Meta Model to Rule Them All
Recommenders
as
features
Machine
learning
to
op5mize
purchasing
probability
Combine
Recommend
Describe

10. Recommender system for Home Page Ordering
Cleaning, combining
and enrichment of
data
Recommendation
Engines
Optimization of
home display
the application
automatically runs and
compiles heterogeneous
data
Generation of
recommendations based
on user behaviour
Every customer is
shown the 10
sales he is the
most likely to buy
Customer visits
Purchases
Sales Images
Metal model combine
recommendations to
directly optimize
purchasing probability
Meta Model
+7% revenue
Sales information
(A/B testing)
Batch Scoring every night

11. Why use Image ?
We want do distinguish
« Sun
and
Beach »
« Ski »
A picture is worth a thousand words

12. Integrating Image Information
Sales Images Labelling
Model
Pool + Palm Trees
Hotel
+ Mountains
Pool + Forest + Hotel
+ Sea
Sea + Beach +Forest +
Hotel
Sales
descriptions
vector
CONTENT
BASED
Recommender
System

13. Image Labelling For Recommendation Engine
Pragma&c
Deep
learning
for
“Dummies”

14. Using Deep Learning models
Common Issues
“I don’t have GPUs server” “I don’t have a deep leaning expert”
“I don’t have labelled data”
(or too few)
“I don’t have the time to wait for model training ”
I don’t want to pay to pay for private apis” / “I’m afraid their labelling will change over time”

15. “I don’t have (or few) labelled data”
-> Is there similar data ?
Solution 1 : Pre trained models
PLACES
DATABASE
US
SUN
DATABASE
205
categories
2.5
M
images
307
categories
110
K
images

16. tower: 0.53
skyscraper: 0.26
swimming_pool/outdoor: 0.65
inn/outdoor: 0.06
Solution 1 : Pre trained models
If there is open data, there is an open pre trained model !
• Kudos to the community
• Check the licensing
Example
with
Places
(Caffe
Model
Zoo)
:

17. Solution 2 : Transfer Learning
Credit
:
Fei-­‐Fei
Li
&
Andrej
Karpathy
&
Jus5n
Johnson
hYp://cs231n.stanford.edu/slides/winter1516_lecture11.pdf

18. Solution 2 : Transfer Learning
Use the network as a feature extractor
Fine tune the model Keras makes it super easy !
(Possible to freeze layer to train just the top) hYps://github.com/PGu5/Deep_Learning_tuto

19. PLACES
DATABASE
US
SUN
DATABASE
Training
(op5onal)
Pre-­‐trained
model
VGG16
tower: 0.53
skyscraper: 0.26
Re-­‐Training
Transferred
Data
:
Last
convolu5onal
layer
features
Re-­‐trained
model
TensorFlow
2
fully
connected
layers
Caffe
Model
Zoo
GPU
CPU
GPU
Leverage existing knowledge !
Solution 2 : Transfer Learning
Accuracy:
72%,
Top-­‐5
Acc:
90
%
>
state
of
the
art
on
dataset
alone
Transfer

20. Post Treatment & Results
(Or how we transfer the labelling information)
Using
Images
informa&on
for
BI
on
steroids

21. Labels post-processing
Complementary information Redondant information
Issue with our approach:
Solution : Matrix Factorization

22. Topic extraction with Non-Negative Matrix Factorization
• Non Negative Matrix factorization (NMF)
X = WH
• X : image x tags, non negative
• W : image x theme
• H : theme x tag
(scikit learn implementation)
• Most represented Themes
• Swimming-pool_Apartment_Putting-green
• Ocean_Coast_SandBar
• Coast_SeaCliff_RockArch
• Beach_Coast_BoardWalk
• Bridge_Viaduc_River
• Palace_BuildingFacade-Mansion
• Castle_Mansion_Monastery
• HotelRoom_Bedroom_DormRoom
• Dimension Reduction
• 200x200 pixels -> 600 tags => 30
themes
• Faster content based filtering
• Image often sparse combination of
themes
Faster content based filtering
• Each theme has the same explication
power
Balanced vector for content based
• Explicability
Each theme corresponds to a few labels

23. Image content detection
Topic scores determine the importance of topics in an image
TOPIC
TOPIC
SCORE
(%)
Golf
course
–
Fairway
–
PuHng
green
31
Hotel
–
Inn
–
Apartment
building
outdoor
30
Swimming
pool
–
Lido
Deck
–
Hot
tub
outdoor
22
Beach
–
Coast
-­‐
Harbor
17
TOPIC
TOPIC
SCORE
(%)
Tower
–
Skyscraper
–
Office
building
62
Bridge
–
River
–
Viaduct
38

24. Note on model performance
• Images labels are used for similarity
Calling herb field “putting green”:
• Is not important if all herbs field are called this way.
• Would be if we had lot’s of golf trips sales.
• Improving the NN performance ?
• Labels are used in NMF and reduced to themes
• Themes are used to calculate similarities for CB
recommenders
• CB Recommenders are used as a feature in meta model
• Meta model give probabilities of purchase = order
• Users only check 10 sales…
-> what is the change of online performance for 1% accuracy ?

25. Results

26. Results ?
1) Visits :
• France and Morocco
• Pool displayed
2) First Recommendation
• Mostly France &
Mediterranean
• Fails to display pools
3) Only Images recommendation
• Pool all around the world
• Does not respect budget
4) Third column = Right Mix

27. Results ?
1) Visits :
• Spain
• Sun & Beach
• Pool displayed
2) First Recommendation
• Displays nature…
3) Only Images ?
• Pool all around the world
4) Third = Right Mix
• Get the bungalow
feature !

28. Learned along the way
What’s next ?
AYrac5veness
=
%
visits
with
tag
/
%
sales
with
tag
-­‐1
For
ski
sales,
indoor
pictures
performs
beYer!

29. What’s Next ?
Kenya
Prague
Berlin
Cambodia

30. Conclusion
Do iterative data science !
Start simple and grow
Evaluate at each steps
Image labelling = BI on steroids
Transfer Learning
Kick-start your project
Gain time and money
Any Data Scientist can do it
Deep Learning
Don’t start from scratch !
Is there existing data ?
Is there a pre-trained model ?

31. Thank you for your attention !

32. Basic Recommendation Engines
• Implementation
• Everything in SQL
- Vertica
- Then Impala
- Then Spark
• Collaborative filtering
• Score(user, future sale j) = Sum_{i:visited sale} sim(i,j)
• Content based filtering
• Sale profile: (sale_id, feature, value)
• User profile: (user_id, feature, value)
• Sparse coding + join does the trick
Join
on
user_id
Join
on
user_id
+
sales
id

33. One Meta Model to Rule Them All
• Negative sampling
• Take all purchases tuples : (user, product, timestamp)-> 1
• Select 5 sales open at the same date the user did not buy -> 0
• The model directly optimize purchasing probability
• Machine learning model
• Features : recommender systems.
• Logistic Regression
Regularizing effect : we don’t want to overfit leaks.
• Reranking approach.
Similar to Google or Yandex (Kaggle challenge)

34. Classification and not Detection problem…
• Only have probabilities of each class
• Selecting based on probability threshold fails
• Keeping all information is not sparse
Labels post-processing
Deep/Transfer Learning
5-10 tags per images2s/image with CPU
x20 speed up with GPU
Our images
Cafe
-> we keep 5 labels
With probabilities per image

35. Solution 3 : What about APIs ?

36. Solution 3 : What about APIs ?
• Price
• Their cost
often rather cheap. Ex: 100 K request for less than 300$
• VS the one of redeveloping (probably not as well)
• Full Database scoring
• APIs are often limited query per month.
• Make sure to be able to avoid cold start problem
• Stability
• Use model versioning
• Avoid covariate shift, distribution drift

37. What about APIs ? Use for generating labels !
How to steal model:
• 1) Score part of the database for training
• 2) Train a model
• 3) Score your entire database !
(Or don’t, it’s illegal)
But I have only 5000 requests ?
-> Use Transfer Learning !
Tramèr, Florian, et al. "Stealing Machine Learning Models via Prediction APIs." arXiv
preprint arXiv:1609.02943 (2016).

38. What about APIs ? Use for generating labels !
Experiment:
• 5000 requests on API. (4500/500 split)
• Transfer learning with MIT Places Pre-trained Model
• Scikit learn Multilabel model
(Or don’t, it’s illegal)
(demo, not used in any real project)

39. What about APIs ? Results
Precision
75
Label
Probability
Label
Probability
landscape 1,0000 sunset 0,9998
sky 1,0000 no person 0,9996
outdoors 1,0000 water 0,9990
nature 1,0000 park 0,9849
rock 1,0000 river 0,9678
travel 1,0000 scenic 0,8031
Label
Probability
Label
Probability
beach 1,0000 ocean 1,0000
summer 1,0000 relaxation 1,0000
sand 1,0000 island 1,0000
tropical 1,0000 idyllic 1,0000
travel 1,0000 seashore 0,9998
seascape 1,0000 water 0,9997
(demo, not used in any real project)
Recall
80
Accuracy
95