Está en la página 1de 37

WHERE

ARE PEOPLE MEETING?

EHARMONY - WHO WE ARE

SOFTWARE ENGINEERS

DATA SCIENTISTS

PSYCHOLOGISTS

PRODUCT SPECIALISTS

MARKETING

CUSTOMER CARE

COMPATIBILITY MATCHING SYSTEM

Anity 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

Anity Matching

Match Distribu2on

AFFINITY MATCHING

Incompa@ble lifestyles

Incompa@ble loca@ons

Very compa@ble - communica@on happens

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 dened by

WIDTH =
HEIGHT 4.038

Face Ra2o is dened 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

Anity 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

WHOS MARRYING ONLINE?

THE BOTTOM LINE

EHARMONY - THE BIG PICTURE