20
Lessons learned in display advertising Natural experiments at scale Robert Moakler ([email protected]) Ekaterina Eliseeva ([email protected]) Kiril Tsemekhman ([email protected]) CODE@MIT 2015

Natural Experiments at Scale

Embed Size (px)

Citation preview

Page 1: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

Lessons learned in display advertising Natural experiments at scale

Robert Moakler – ([email protected]) Ekaterina Eliseeva – ([email protected]) Kiril Tsemekhman – ([email protected])

CODE@MIT 2015

Page 2: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

The $100+ billion question!

Does online advertising really work?

$104.57 $120.05

$140.15 $160.18

$178.45 $196.05

$213.89 Digital ad spending!% change!

2012 2013 2014 2015 2016 2017 2018!Source: www.emarketer.com, “Global Ad Spending Growth to Double This Year”

20.4% 14.8% 16.7% 14.3% 11.4% 9.9% 9.1%

Page 3: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

The $100+ billion question!

Does online advertising really work?

Do online ads cause you to take some action?

Page 4: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

The usual approach!Randomized experiments and A/B tests are great!

Campaign Ad PSA

Page 5: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

The usual approach!But sometimes …

RIGHT WRONG

Randomized experiments and A/B tests are great! Campaign Ad PSA

Page 6: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

Natural experiments!•  Consider the typical setup for the ad serving process

Confounding!

W User

features

A Served

ads

Y Convert

Page 7: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

Natural experiments!•  Consider the typical setup for the ad serving process •  Introduce a mediating variable

W User

features

A Served

ads

Y Convert

M Mediator

W’ Residual

Confounders

Page 8: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

Natural experiments!•  Consider the typical setup for the ad serving process •  Introduce a mediating variable

–  Viewability

W User

features

A Served

ads

Y Convert

V Viewable

ad

W’ Residual

Confounders

Page 9: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

Ad viewability!

Horizontal location (px) Proportion of ads

Verti

cal l

ocat

ion

(px)

Ad density

Page 10: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

Running in the wild!•  Natural experiments aren’t always clean or easy

•  We will discuss five problems that we have run into and some solutions for dealing with them

Page 11: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

An online advertising campaign!•  Our data structure

Analysis window

Viewable ad Unviewable ad Conversion Web activity

Our users

Page 12: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

Longitudinal data!•  Monitoring

–  Most online advertising campaigns run continually –  We are constantly monitoring many campaigns at the event level

•  Running an intermediary analysis –  Data is subject to left truncation and right censoring –  We need to account for our residual confounders, W’

•  Use survival analysis –  Cox Proportional Hazards (CPH) model

Page 13: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

User fragmentation and study period!•  In reality, our users are defined by cookies.

–  However, people do not just have one cookie!

Viewable ad Unviewable ad Conversion Web activity

Sarah Cookie 1 Cookie 2 Cookie 3

Bob Cookie 1 Cookie 2

Analysis window

Cookie 3

Page 14: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

User fragmentation and study period!•  In reality, our users are defined by cookies.

–  However, people do not just have one cookie!

•  Some methods we use to account for this –  We define an effect period of 1 week

•  Seasonality has a major impact •  Users are selected through iterative simulation and research •  Incremental causal estimates level off after a single week

Page 15: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

Validation!•  How do we know our causal models give reasonable estimates? •  Use an array of negative control tests

–  Use the impressions of one campaign to predict an unrelated conversion

W User

features

A Served

ads

Y Convert

W’ Residual

Confounders

Y Unrelated

Event

-

V Viewable

ad

Page 16: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

Running at scale!•  Converting our data into something analyzable is a challenge

Raw daily logs Billions of events

HDFS scalable cluster storage

Hadoop

People browse the web.

Advertising events turn into billions of daily events.

Raw data is moved to scalable storage optimized for our experimental setup.

Users are subsampled and negative controls are chosen in parallel.

Reports are run in parallel using stripped down R libraries.

Iterative process of simulation and

research.

Page 17: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

Summary!•  Mediators and natural experiments may already exist in your data

•  Running a natural experiment at scale is not straight forward, because 1.  The longitudinal nature of the data 2.  Users can become highly fragmented 3.  No predetermined start and end dates 4.  Validation of causal models 5.  Billions of events and terabytes of raw data

•  Equal parts engineering and modeling

•  We explored online advertising, but this setup can apply to a wide variety of industries

Page 18: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

Thanks! Robert Moakler – ([email protected]) Ekaterina Eliseeva – ([email protected]) Kiril Tsemekhman – ([email protected]) Grab this deck @ bit.ly/natural-experiments-at-scale

Page 19: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

Acknowledgments!Integral Ad Science Ekaterina Eliseeva Kiril Tsemekhman Ana Calabrese Gijs Joost Brouwer Sergei Izrailev

NYU Stern Foster Provost Amazon, Inc. Daniel Hill

Page 20: Natural Experiments at Scale

CODE@MIT – OCTOBER 16 2015

References!Chan, D., Ge, R., Gershony, O., Hesterberg, T., & Lambert, D. (2010, July). Evaluating online ad

campaigns in a pipeline: causal models at scale. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 7-16). ACM.

Dalessandro, B., Perlich, C., Stitelman, O., & Provost, F. (2012, August). Causally motivated attribution for online advertising. In Proceedings of the Sixth International Workshop on Data Mining for Online Advertising and Internet Economy (p. 7). ACM.

Hill, D. N., Moakler, R., Hubbard, A. E., Tsemekhman, V., Provost, F., & Tsemekhman, K. (2015, August). Measuring Causal Impact of Online Actions via Natural Experiments: Application to Display Advertising. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 1839-1847). ACM.

Johnson, G. A., Lewis, R. A., Nubbemeyer, E. I. (2015, October). Ghost Ads: Improving the Economics of Measuring Ad Effectiveness. Available on SSRN: ssrn.com/abstract=2620078

Klein, J. P., & Moeschberger, M. L. (2003). Survival analysis: techniques for censored and truncated data. Springer Science & Business Media.

Pearl, J. (2009). Causality. Cambridge university press.