46
GROWTH HACKING AND THE BANDIT PROBLEM Starring Noel Welsh A Production Friday, 8 February 2013

Growth Hacking and the Bandit Problem

Embed Size (px)

DESCRIPTION

Growth hackers rely on A/B testing to validate their improvement, but A/B testing is slow and inefficient. How do we make it faster? By solving the bandit problem! Bandit algorithms provide an alternative to A/B testing that is faster and more efficient, allowing more tests in less time. Video of the talk: http://www.youtube.com/watch?v=h7SG_Wo8sPQ Myna: http://mynaweb.com/?utm_source=awareness&utm_medium=slideshare&utm_campaign=Growth%2BHacker

Citation preview

Page 1: Growth Hacking and the Bandit Problem

GROWTH HACKING AND

THE BANDIT PROBLEM

Starring Noel Welsh A Production

Friday, 8 February 2013

Page 2: Growth Hacking and the Bandit Problem

WHAT DO

GROWTH HACKERS DO?

Friday, 8 February 2013

Page 3: Growth Hacking and the Bandit Problem

GROW!Friday, 8 February 2013

Page 4: Growth Hacking and the Bandit Problem

Metric(sign-ups, revenue, etc)

Friday, 8 February 2013

Page 5: Growth Hacking and the Bandit Problem

Make this happen

Friday, 8 February 2013

Page 6: Growth Hacking and the Bandit Problem

HOW?Friday, 8 February 2013

Page 7: Growth Hacking and the Bandit Problem

Build

MeasureLearn

Friday, 8 February 2013

Page 8: Growth Hacking and the Bandit Problem

Build

Friday, 8 February 2013

Page 9: Growth Hacking and the Bandit Problem

Build

Measure

Friday, 8 February 2013

Page 10: Growth Hacking and the Bandit Problem

Build

MeasureLearn

Friday, 8 February 2013

Page 11: Growth Hacking and the Bandit Problem

FASTER IS BETTER

Friday, 8 February 2013

Page 12: Growth Hacking and the Bandit Problem

HOWDO WE MEASURE &

LEARN?

Friday, 8 February 2013

Page 13: Growth Hacking and the Bandit Problem

A/BFriday, 8 February 2013

Page 14: Growth Hacking and the Bandit Problem

MeasureBuild Learn

Friday, 8 February 2013

Page 15: Growth Hacking and the Bandit Problem

Build

MeasureLearn

Friday, 8 February 2013

Page 16: Growth Hacking and the Bandit Problem

Build

Measure/Learn

Friday, 8 February 2013

Page 17: Growth Hacking and the Bandit Problem

Build Measure / Learn

Friday, 8 February 2013

Page 18: Growth Hacking and the Bandit Problem

THE BANDIT PROBLEM(EST 1952 ... OR 1933)

Friday, 8 February 2013

Page 19: Growth Hacking and the Bandit Problem

k arms (or variants)

Friday, 8 February 2013

Page 20: Growth Hacking and the Bandit Problem

Select a variant

Friday, 8 February 2013

Page 21: Growth Hacking and the Bandit Problem

Receive a reward

Friday, 8 February 2013

Page 22: Growth Hacking and the Bandit Problem

REPEAT

Friday, 8 February 2013

Page 23: Growth Hacking and the Bandit Problem

MAXIMISE TOTAL REWARD

Friday, 8 February 2013

Page 24: Growth Hacking and the Bandit Problem

Friday, 8 February 2013

Page 25: Growth Hacking and the Bandit Problem

HOW?Friday, 8 February 2013

Page 26: Growth Hacking and the Bandit Problem

DISPLAY VARIANT WITH

HIGHEST AVERAGE REWARD

Friday, 8 February 2013

Page 27: Growth Hacking and the Bandit Problem

Good Variant Bad Variant

Friday, 8 February 2013

Page 28: Growth Hacking and the Bandit Problem

Good Variant Bad Variant

Views: 1Reward: 0Average: 0

Friday, 8 February 2013

Page 29: Growth Hacking and the Bandit Problem

Good Variant Bad Variant

Views: 1Reward: 0Average: 0

Views: 1Reward: 1Average: 1

Friday, 8 February 2013

Page 30: Growth Hacking and the Bandit Problem

OOPS!Friday, 8 February 2013

Page 31: Growth Hacking and the Bandit Problem

Good Variant Bad Variant

Views: 1Reward: 0Average: 0

Views: 1Reward: 1Average: 1

WILL NEVER TRY THE

GOOD VARIANT AGAIN!

Friday, 8 February 2013

Page 32: Growth Hacking and the Bandit Problem

BALANCE EXPLOIT

AND EXPLORE

Friday, 8 February 2013

Page 33: Growth Hacking and the Bandit Problem

EXPLOITTRY THE VARIANT THAT HAS WORKED BEST IN

THE PAST

Friday, 8 February 2013

Page 34: Growth Hacking and the Bandit Problem

EXPLORE TRY OTHER VARIANTS

TO SEE IF THEY’RE BETTER

Friday, 8 February 2013

Page 35: Growth Hacking and the Bandit Problem

E-GREEDYFriday, 8 February 2013

Page 36: Growth Hacking and the Bandit Problem

10% EXPLORE

90% EXPLOITDon’t like 10% and 90%? Change it.

Friday, 8 February 2013

Page 37: Growth Hacking and the Bandit Problem

MORE INTRICATE BALANCING IS POSSIBLE.

SEE UCB-1, THOMPSON SAMPLING, etc.

Friday, 8 February 2013

Page 38: Growth Hacking and the Bandit Problem

Friday, 8 February 2013

Page 39: Growth Hacking and the Bandit Problem

DOES IT

WORK?

Friday, 8 February 2013

Page 40: Growth Hacking and the Bandit Problem

Friday, 8 February 2013

Page 41: Growth Hacking and the Bandit Problem

REAL WORLD

CONSIDERATIONS

Friday, 8 February 2013

Page 42: Growth Hacking and the Bandit Problem

WORKFLOW

Friday, 8 February 2013

Page 43: Growth Hacking and the Bandit Problem

DEFINING

REWARDS

Friday, 8 February 2013

Page 44: Growth Hacking and the Bandit Problem

STABLE PREFERENCES

Friday, 8 February 2013

Page 45: Growth Hacking and the Bandit Problem

THANK YOU!NOW GO FORTH AND

ENGAGE IN BANDITRY

Friday, 8 February 2013

Page 46: Growth Hacking and the Bandit Problem

CREDITSPhotos and artwork by the following people were used in

this presentation:

Moyan Brenn http://www.flickr.com/photos/aigle_dore/Suicine http://www.flickr.com/photos/bigmikeyeah

ankakay http://www.flickr.com/photos/ankakay/Cdr Aitch http://www.flickr.com/people/hjsouthgate/

StJost http://stjost.deviantart.com/

Friday, 8 February 2013