response prediction for display advertising - wsdm 2014

RESPONSE PREDICTION FORDISPLAY ADVERTISING

WSDM ’14

Olivier Chapelle

2

OUTLINE

1. Criteo & Display advertising

2. Modeling

3. Large scale learning

4. Explore / exploit

3

DISPLAY ADVERTISING

Display Ad

4

DISPLAY ADVERTISING

• Rapidly growing multi-billion dollar business (30% of internet advertising revenue in 2013).

• Marketplace between:– Publishers: sell display opportunities– Advertisers: pay for showing their ad

• Real Time Bidding:– Auction amongst advertisers is held at the moment when a user generates

a display opportunity by visiting a publisher’s web page.

5

BRANDING VS PERFORMANCE

PRICING TYPE

CAMPAIGN TYPE

• CPM (Cost Per Mille): advertiser pays per thousand impressions

• CPC (Cost Per Click): advertiser pays only when the user clicks

• CPA (Cost Per Action): advertiser pays only when the user performs a predefined action such as a purchase.

• Branding CPM

• Performance based advertising (retargeting) CPC, CPA

CONVERSIONS

eCPM = CPC * predicted clickthrough rate

eCPM = CPA * predicted conversion rate

6

CRITEO

6

Bill on

CPC

Advertisers

Pay on

CPM

Publishers

Criteo’s success depends highly on how precisely we can predict Click Through Rates (CTR) and Conversion Rates (CR)

7

RECOMMENDATION

Collaborative filteringPropose related items to a user based on historical interactions from other users

Over 50% of Criteo driven sales come from recommended products the user had never viewed on advertiser websites

8

CRITEO NUMBERS

8

750EMPLOYEES

K250AD CALLSPER SECOND

107500 CORES CLUSTER

PBB

AN

NE

RS

PE

R D

AY 2.1

BILLION

BILLION18

RT

BID

SP

ER

DA

Y

42COUNTRIES

$600MILLION2013 REVENUE

9

OUTLINE

1. Criteo & Display advertising

2. Modeling

3. Large scale learning

4. Explore / exploit

10

• Three sources of features: user, ad, page• In this talk: categorical features on ad and page.

FEATURES

Publisher network

Publisher

Site

Url

Advertiser network

Ad

Campaign

Advertiser

Publisher hierarchy Advertiser hierarchy

11

HASHING TRICK

• Standard representation of categorical features: “one-hot” encoding

For instance, site feature

• Dimensionality equal to the number of different values– can be very large

• Hashing to reduce dimensionality (made popular by John Langford in VW)

• Dimensionality now independent of number of values

cnn.com news.yahoo.com

0 0 01 0 0 0

12

HASHING VS FEATURE SELECTION

• “Small” problem with 35M different values.• Methods that require a dictionary have a larger model.

13

QUADRATIC FEATURES

• Outer product between two features.• Example: between site and advertiser,

Feature is 1 site=finance.yahoo.com & advertiser=bank of america

Similar to a polynomial kernel of degree 2 Large number of values hashing trick

Publisher network

Publisher

Site

Url

Advertiser network

Ad

Campaign

Advertiser

14

ADVANTAGES OF HASHING

• Practical– Straightforward implement; no need to maintain dictionaries

• Statistical– Regularization (infrequent values are washed away by frequent ones)

• Most powerful when combined with quadratic features

Quote of John Langford about hashing

At first it’s scary, then you love it

15

LEARNING

• Regularized logistic regression– Vowpal Wabbit open source package

• Regularization with hierarchical features backoff smoothing

• Negative data subsampled for computational reason

Well estimated Small if rare value

16

EVALUATION

• Comparison with (Agarwal et al. ’10)– Probabilistic model for the same display advertising prediction problem – Leverages the hierarchical structures on the ad and publisher sides– Sparse prior for smoothing

• Model trained on three weeks of data, tested on the 3 following days

auROC auPRC Log likelihood

+ 3.1% + 10.0% + 7.1%

D. Agarwal et al., Estimating Rates of Rare Events with Multiple Hierarchies through Scalable Log-linear Models, KDD, 2010

17

BAYESIAN LOGISTIC REGRESSION

• Regularized logistic regression = MAP solution (Gaussian prior, logistic likelihood)

• Posterior is not Gaussian• Diagonal Laplace approximation:

with:

and:

18

MODEL UPDATE

• Needed because ads / campaigns keep changing.• The posterior distribution of a previously trained model can be used as

the prior for training a new model with a new batch of data

• Influence of the update frequency (auPRC):

1 day 6 hours 2 hours

+3.7% +5.1% +5.8%

Day 1 Day 2 Day 3 Day 4 Day 5

M0 M1 M2

19

OUTLINE

1. Criteo & Display advertising

2. Modeling

3. Large scale learning

4. Explore / exploit

20

PARALLEL LEARNING

• Large training set– 2B training samples; 16M parameters– 400GB (compressed)

• Proposed method: less than one hour with 500 machines

• Optimize:

• SGD is fast on a single machine, but difficult to parallelize.• Batch (quasi-Newton) methods are straightforward to parallelize

– L-BFGS with distributed gradient computation.

= + … +S S1 Sm

21

ALLREDUCE

• Aggregate and broadcast across nodes

• Very few modification to existing code: just insert several AllReduce op.• Compatible with Hadoop / MapReduce

– Build a spanning tree on the gateway– Single MapReduce job– Leverage speculative execution to alleviate the slow node issue

22

ONLINE INITIALIZATION

• Hybrid approach:– One pass of online learning on each node– Average the weights from each node to get a warm start for batch

optimization

• Best of both (online / batch) worlds.

Splice site prediction (Sonnenburg et al. ‘10) Display advertising

S. Sonnenburg and V. Franc, COFFIN: A Computational Framework for Linear SVMs, ICML 2010

23

OUTLINE

1. Criteo & Display advertising

2. Modeling

3. Large scale learning

4. Explore / exploit

24

THOMPSON SAMPLING

• Heuristic to address the Explore / Exploit problem, dating back to Thompson (1933)

• Simple to implement• Good performance in practice (Graepel et al. ‘10, Chapelle and Li ‘11)• Rarely used, maybe because of lack of theoretical guarantee.

T. Graepel et al., Web-scale Bayesian click-through rate prediction for sponsored search advertising in Microsoft’s Bing search engine, ICML 2010

O. Chapelle and L. Li, An Empirical Evaluation of Thompson Sampling, NIPS 2011

Draw model parameter qt according to P( | q D)

Select best action according to qt

Observe reward and update model

25

E/E SIMULATIONS

MAB with K arms. Best arm has mean reward = 0.5, others have 0.5 − ε.

26

EVALUATION

• Semi-simulated environment: real input features, but labels generated.• Set of eligible ads varies from 1 to 5,910. Total ads = 66,373• Comparison of E/E algorithms:

– 4 days of data– Cold start

• Algorithms:– UCB: mean + a std. dev. – e-greedy– Thompson sampling

27

RESULTS

• CTR regret (in percentage):

• Regret over time:

Thompson UCB e-greedy Exploit-only Random

3.72 4.14 4.98 5.00 31.95

28

OPEN QUESTIONS

• Hashing– Theoretical performance guarantees

• Low rank matrix factorization– Better predict on unseen pairs (publisher, advertiser)

• Sample selection bias– System is trained only on selected ads, but all ads are scored.– Possible solution: inverse propensity scoring– But we still need to bias the training data toward good ads.

• Explore / exploit– Evaluation framework– Regret analysis of Thompson’s sampling– E/E with a budget; with multiple slots; with a delayed feedback

29

CONCLUSION

• Simple yet efficient techniques for click prediction

• Main difficulty in applied machine learning: avoid the bias (because of academic papers) toward complex systems– It’s easy to get lured into building a complex system– It’s difficult to keep it simple

See paper for more details

Simple and scalable response prediction for display advertising

O. Chapelle, E. Manavoglu, R. Rosales, 2014