which car fits my life? - inovex · deepstack: expert-level artificial intelligence in heads-up...

Which car fits my life?Mobile.de’s approach to recommendations

PyData Berlin, July 1st, 2017Florian Wilhelm, Arnab Dutta

2

Introduction

Dr. Florian Wilhelm

Data Scientist

inovex GmbH

� @FlorianWilhelm

� FlorianWilhelm

florianwilhelm.info

Dr. Arnab Dutta

Data Scientist

mobile.de GmbH

� @kopfhohen

� kraktoso

3

Outline

§Introduction§Use-cases§Theory§Our Approach§Example§Outlook

4

MOBILE.DEGERMAN MARKET LEADER

13.5 MIO UNIQUE USER PER MONTH

1.6 MIO VEHICLES

290EMPLOYEES

DREILINDEN / FRIEDRICHSHAIN BERLINHEADQUARTERS

Part ofebay Tech

5

IT-project house for digital transformation:‣ Agile Development & Management‣ Web · UI/UX · Replatforming · Microservices‣ Mobile · Apps · Smart Devices · Robotics‣ Big Data & Business Intelligence Platforms‣ Data Science · Data Products · Search · Deep Learning‣ Data Center Automation · DevOps · Cloud · Hosting‣ Trainings & Coachings

Using technology to inspire our clients. And ourselves.

inovex offices inKarlsruhe · Pforzheim · München · Köln · Hamburg.

www.inovex.de

6

Outline

§Introduction§Use-cases§Theory§Our Approach§Example§Outlook

7

Why Recommendations?Why Recommendations?

Show width of offering

Inspiration

Engagement

8

- - engagement- - inspiration- - relevance

Why Recommendations?

- - high click-through-rate - - small exit- & bounce-rates

User Benefits

Business Benefits

9

Mobile.de Conversion Funnel

WishlistHome Search Result Page

View Contact Buy

10

Recommendations on Home

Home

Recommendations based on preferencesof visiting users as an alternative entry point.

WishlistHome SRP View Contact Buy

11

Recommendations on Search Results Page

Recommendations basedon similar vehicle makeand model id to presentalternatives

WishlistHome SRP View Contact Buy

12

Recommendations on View Item Page

VIP

Recommendations based on the specific make and model a user is viewing to present alternatives

WishlistHome SRP View Contact Buy

13

Recommendations on your Wishlist

Recommendations based on the specific make and model of a deleted ad to provide almost identical recommendations

Recommendations based on theusers car preferences and the parking lot items.

WishlistHome SRP View Contact Buy

14

§Introduction§Use-cases§Theory§Collaborative Filtering§Content Based§Our Approach§Example§Outlook

15

Collaborative Filtering

items similar

16

Watched / Rated

Unwatched

Item-Item Similarity

??

??

????Item-based Recommendations

Cosine Similarity

17

Recommendation

Item-based Recommendations

P

P

P

P

P P

Wishlist

18

Summary of Collaborative Filtering✓Collective behaviour of users

✓Standard-Method (it works, it’s reliable etc.)

� Cold Start Problem: New listings need a certain number of clicks to be recommended.

� Sparsity problems: lot fewer interactiondata points than total items and users.

� Content agnostic

� Only “batch-based” learning

19

Looking For: Used Car (100%)

Prefers (Make): BMW (50%), Audi (50%)Prefers (Model): Audi A3 (25%), Audi A4 (25%),

BMW 318 (50%)Searching In: lat 52.5206, lon 13.409Search Radius: 300kmPreferred Price: 20 000€ ± 1500€ Preferred Mileage: 10 000km ± 5000km

User Preferences

Marketing

Anonymous

Content-based Filtering: User Preferences

20

Content-based Filtering

interacted

<Price: 10K, Category: small>

<Price: 6K, Category: small>

<Price: 90K, Category: sports>

<Price: 10K, Category: small>

recommend

21

Summary of Content-basedüWorks even if there are no other

users

ücontent-based preferences ofusers based on a weighted vectorof item features

� Hard to do recommendations fornew users (cold start problem)

� Non-applicable for heterogenouscontent types

� Low diversity, i.e. more of thesame

22

§Introduction§Use-cases§Theory§Our Approach§Example§Outlook

23

Find the car that perfectly fits your life

User’s Car Preferences Car Pool + Attributes(make, model, color, price, …)

Flexible(cold-start, uncertainty, real-time, ...)

Interactions of other users(views, parkings, contacts)

24

Hybrid Recommender

Collaborative Filtering

Hybrid Recommender

Contentbased

• no cold-start problem of new items

• integrate new user events in real-time

• robust and reliable concepts

• easy to tune for different use-cases

• comprehensible and debuggable

25

RecommendationEngine

User PreferenceService API

User EventTracking & Storage

RecommendationEngine

RecommendationEngine

User PreferenceComputation &

Storage

All Listings

User Preference+ Recommendation Architecture

26

Hybrid Recommender Concept

PPP P

P

Looking For: Used Car (100%)

Prefers (Make): BMW (50%), Audi (50%)Prefers (Model): Audi A3 (25%), Audi A4 (25%),

BMW 318 (50%)

Searching In: lat 52.5206, lon 13.409Search Radius: 300kmPreferred Price: 20 000€ ± 1500€Preferred Mileage: 10 000km ± 5000km

User Profile

BuyerLast Action: YesterdayFrequent User

User 12345

Likelihood to buy: 88 %

Elastic Search Query

Score0.8

0.3

1.7

3.2

1.1

0.9

……

27

§Introduction§Use-cases§Theory§Our Approach§Example§Outlook

28

Finding similar make/models

§Users are often uncertainwith their choices in orientation phase

§Help users explore similarmodels to make informeddecisions

§Exploit the collaborativeaspect in defining theconcept of similarity

29

Make/Model Recommender with LightFM

LightFM:§Matured and well documented Python package§Optimized and parallelized with Cython§Hybrid recommender based on matrix factorisation§Supports Learning-to-Rank objectives (BPR, WARP)

Audi A4 BMW 645 similar vehicles

30

Non-negative Matrix Factorisation (NMF)LF1 LF2

LF1

LF2

M (|U| x |I|) x R (|LF| x |I|)

= X

= L (|U| x |LF|)

31

NMF: Embeddings to Similarities

LF1

LF2

• car is represented as an item embedding

• Given 2 embeddings, LFitem_i and Lfitem_j_

• Compute sim(LFitem_i , LFitem_j)

• Find pairwise values for all item pairs (|I| x |I|)

item embeddings

32

NMF: Persisting Item similarities

33

Method§ best of both the worlds§ robust

Business§ higher CTR§ lesser exits rates

User§ engagement§ diversity

hybrid

34

§Introduction§Motivation & Use-cases§Theory§Implementation§Outlook

35

Deep Learning

Dermatologist-level classification of skin cancer with deep neural networks(Nature 542, 115-118, February 2017)

DeepStack: Expert-level artificial intelligence in heads-up no-limit poker(Science, March 2017)

Recent Breakthroughs in Deep Learning Reasons for Deep Learning

• captures nonlinear relations

• holistic approach, i.e. reduces number of components possibly

• less feature engineering

• possibly improved quality

36

Approach: Wide and Deep Model

mileagepricecolor

history

mileagepricecolor

views

...

0.38

0.250.79

...0

10

...0

0.20.8

...

...

enco

deen

code

enco

deen

code

...

0.35

-0.152.03

cont.

cat.

cont.

cat.us

erem

bedd

ings

item

em

bedd

ings

Outp

ut

cros

s-ite

m

Probability that user X likes vehicle Y

Deep Component

Wide Component

one-

hot

37

Recommendations support users to find the perfect vehicle based on their preferences and by collaboration

38

Any questions?lusion