WHERE  ARE  PEOPLE  MEETING?

EHARMONY  -­‐  WHO  WE  ARE

SOFTWARE  ENGINEERS

DATA  SCIENTISTS

PSYCHOLOGISTS

PRODUCT  SPECIALISTS

MARKETING

CUSTOMER  CARE

COMPATIBILITY  MATCHING  SYSTEM

Affinity  Matching

Match  Distribu2on

COMPATIBILITY  MATCHING  SYSTEM

COMPATIBILITY  MATCHING  SYSTEM

COMPATIBILITY  MATCHING  SYSTEM

COMPATIBILITY  MATCHING  SYSTEM

COMPATIBILITY  MATCHING  SYSTEM

COMPATIBILITY  MATCHING  SYSTEM

COMPATIBILITY  MATCHING  SYSTEM

COMPATIBILITY  MATCHING  SYSTEM

OUTPUT  
We  are  le7  with  a  number    
of  pairings  for  each  user

COMPATIBILITY  MATCHING  SYSTEM

Affinity  Matching

Match  Distribu2on

AFFINITY  MATCHING

1  

Incompa@ble  lifestyles

2  

Incompa@ble  loca@ons

Very  compa@ble  -­‐  communica@on  happens

3  

Incompa@ble  ages

AFFINITY  MATCHING

FEATURES

PHOTOS  INFORMATION

VARIOUS  PREFERENCES

PROFILE  INFORMATION

SITE  USAGE

FACIAL  DETECTION  -­‐  VIOLA  JONES

P.  Viola,  M.  Jones  Rapid  Object  Detec2on  using  a  Boosted  Cascade  of  Simple  Features.  CVPR  2001

FACIAL  DETECTION  -­‐  VIOLA  JONES

P.  Viola,  M.  Jones  Rapid  Object  Detec2on  using  a  Boosted  Cascade  of  Simple  Features.  CVPR  2001

FACE  DETECTION

IMAGE  FEATURES

Aspect  Ra2o  is  defined  by

WIDTH =
HEIGHT 4.038

Face  Ra2o  is  defined  by

FACE  AREA
=
IMAGE  AREA 0.177

FACE  RATIO

FACE  RATIO

FACE  RATIO

Score = 5.98 + 100.96*x - 1279.95*x^2 + 6483.85*x^3 -14767.15*x^4 + 12492.38*x^5

FACE  DETECTION

FACE  PARTS

App(Image,  Part  Loca@on)

+

Shape(Part  Loca@on)

=

Model(Image,  Part  Loca@on)

hUp://github.com/eharmony/face-­‐parts-­‐service
X.  Zhu,  D.  Ramadan  Face  Detec2on,  Pose  Es2ma2on  and  Landmark  Localiza2on  in  the  Wild.  CVPR  2012

AFFINITY  MATCHING  -­‐  FACIAL  ANALYSIS

AFFINITY  MATCHING  -­‐  NEW  PHOTO  FEATURES

HAIR  COLOR

POSE

EYE  COLOR

HAS  MUSTACHE?
HAS  BEARD?
HAS  CLEAVAGE?

AFFINITY  MATCHING  -­‐  NEW  PHOTO  FEATURES

 AVERAGE  EYE  WIDTH

 CHEEKBONE  WIDTH  
 UPPER  FACE  HEIGHT

AVERAGE  DISTANCE  
FROM  EYE  TO  EYEBROW
CHEEKBONE  WIDTH  
JAW  WIDTH

 FACE  WIDTH  
 LOWER  FACE  HEIGHT
 LOWER  FACE  HEIGHT  
 FACE  HEIGHT

FACE  PERIMETER  
FACE  AREA

L.  Wen,  G.  Guo  A  computa2onal  approach  to  body  mass  index  predic2on  from  face  images.  
Image  and  Vision  Compu2ng,  2013

AFFINITY  PROCESS
~60M  registered  users
~10^7  Matches  per  day
MODEL  THAT  PREDICTS    
COMMUNICATION

~10^3  Features

DSL

20%

45%

38%

50%

75%

82%

COMPATIBILITY  MATCHING  SYSTEM

Compa2bility  Matching

Affinity  Matching

Match  Distribu2on

MATCH  DISTRIBUTION  -­‐  PROBLEM
(1,-­‐A1,1)
(1,-­‐A1,2)

(L1,0)

(LN+1,0)
(LN+2,0)

(L2,0)
(LN,0)

.

.  
.

(1,-­‐A2,M)

(1,-­‐AN,2)
(1,-­‐AN,M)

.

.  
.

(LN+M,0)

MATCH  DISTRIBUTION  -­‐  PROBLEM
(1,-­‐A1,1)
(LN+1,0)

(L1,0)

(LN+2,0)
(L2,0)
(LN,0)

.

.  
.

(1,-­‐A2,M)

(1,-­‐AN,2)

.

.  
.

(LN+M,0)

MATCH  DISTRIBUTION

MATCH  GOAL:  An  amount  of  matches  is  assigned

8  MATCHES

6  MATCHES

Can  we  do  be_er?

10  MATCHES

HISTORY  -­‐  CONTEXTUAL  BANDITS

M.  Dudki,  J.  Langford,  L.  Li,  “Doubly  Robust  Policy  Evalua2on  and  Learning.”  ICML,  2011.

MATCH  GOALS  -­‐  CONTEXTUAL  BANDITS

This  model  is  used  for  training
8  MATCHES

6  MATCHES

10  MATCHES

RESULT:  BETTER  COMMUNICATION

8  MATCHES

6  MATCHES

10  MATCHES

MATCH  GOALS  -­‐  CONTEXTUAL  BANDITS
-0.19

-0.198

-0.205

-0.213

-0.22
Constant 6 Constant 7 Constant 8 Constant 9 Constant 10 Uniform

Model

WHO’S  MARRYING  ONLINE?

THE  BOTTOM  LINE

EHARMONY  -­‐  THE  BIG  PICTURE

Sign up to vote on this title
UsefulNot useful