Asim Ansari

1

E-Customization

• Asim Ansari
• Carl Mela

2
Introduction

• Customization key to managing relationships

• Marketing
• Targeted Promotions
• List Segmentation
• Conjoint Analysis
• Recommendation Systems

• Computer Science
• Collaborative filtering
• Machine learning

3
Customization and Electronic Media

• Electronic media facilitate customization
• Low production costs
• Timely data (received) and information (sent)
• Personalizable
• Reach

4
Customization Benefits

• Content Providers
• Increasing site usage via customization can
  increase advertising revenue
• Internet Advertising forecast to grow to rapidly
• E-commerce
• Increasing sales via customization

5
E-Customization Contexts

• Content providers can customize
• content (editorial)
• design (how many links and what order)
• to increase site visits, advertising revenue and
  loyalty.
• E-commerce firms can customize
• content (products, price, incentives, etc.) and
• design (how many items and what order)
• to increase sales and loyalty.
• The structure of the problem is identical.

6
E-Customization Strategies

• Two customization Strategies
• Onsite
• External e-mails
• Customizable at low-cost
• Need not wait for customers to come to site
• We take an external customization approach

7
E-mail Example
8
E-mail Marketing Volume Growth
250
200
Emails
150
(billions)
100
50
0

99

00

01

02

03

04
Email acquisition services
Source Forrester Report Email Marketing Dialog,
January 2000
Email retention services
9
E-mail Marketing Services Revenue Growth
5
4
Revenues
3
(billions)
2
1
0

99

00

01

02

03

04
Email acquisition services
Source Forrester Report Email Marketing Dialog,
January 2000
Email retention services
10
Email Design Problem
Determine the Content and Layout of the e-mail on
a one-on-one basis
Sports International News National
News Weather Arts
11
Approach
Individual level preference coefficients
E-mail Configuration
Statistical Model
Optimization
New E-mail Configuration
Click-through Data
12
Statistical Model

• Probability of clicking on a link depends upon
  utility to click
• Utility of clicking on a link
• f(
• observed e-mail variables (html, links),
• observed link variables (content and order of
  link),
• unobserved user effect,
• unobserved e-mail effect,
• unobserved link effect,
• error)

13
Probit Model
Population Component

• Uijk m1m2Textjm3NumItemsjm4Positionjk
• m5Contentk
• li1li2NumItemsjli3 Positionjk
  li4Contentk
• qj1qj2Positionjkqj3Contentk
• gk1
• eijk
• i is person, j is e-mail and k is link.

Random across Individuals
Random across Emails
14
Modeling Heterogeneity

• Random effects are assumed to come from a
  population distribution with zero mean
• li G1
• qj G2
• gk G3

15
Modeling Heterogeneity

• Finite Mixtures

• Continuous Mixtures

16
Modeling Heterogeneity Dirichlet Process Priors

• Dirichlet Process Priors can be used to model the
  uncertainty about functional form of the
  population distribution G
• Allows semi-parametric estimation of random
  effects

17
Dirichlet Process Priors

• A Dirichlet Process prior for a distribution G
  has two parameters
• A distribution function G0(.) and
• A positive scalar precision parameter a
• We write
• where, G0 represents the expected value of G and
  a gt 0, represents the strength of prior beliefs
  that sampled distributions G will be close to G0

18
Dirichlet Process Priors

• Let G be a random distribution from the Dirichlet
  Process,
• Let then,

19
Dirichlet Process Role of a

• Large a
• Large number of distinct values from the base
  distribution
• Sampled distribution approximates base
  distribution
• Small a
• Sample will have a small number of distinct
  values
• Sampled distribution approximates a finite mixture

20
Dirichlet Process Priors Advantages

• Accommodates non-normality, multi-modality and
  skewness
• Provides a semi-parametric alternative to the
  normal distribution
• Provides accurate individual-level estimates
• Allows a synthesis of Finite Mixtures and Normal
  Heterogeneity

21
Modeling Heterogeneity

• li G1 D(N(0,L), a1)
• qj G2 D(N(0,Q), a2)
• gk G3 D(N(0,t), a3)

22
Inference

• Bayesian Inference
• Priors
• m Multivariate Normal
• L-1 Wishart
• Qll Inverse Gamma
• t Inverse Gamma
• a1, a2, a3, Gamma

23
Sampling Based Inference

• Joint Posterior Density is very complex and
  cannot be summarized in closed form
• Sampling Based Inference
• Gibbs Sampling

24
Full Conditionals

• Unknowns include
• u, m, li, qj, gk, L, Q, t, a1, a2, a3
• Full conditionals for DP mixed model are very
  similar to those for normal population
  distributions

25
Full Conditionals for Individual-level
parameters DP model

• Mixture of distributions

And Gb is the posterior distribution under the
normal base distribution This is akin to
collaborative filtering on parameter space
26
Application

• Large content provider with many areas in site
• One area in the site sends e-mails to registered
  recipients in an effort to attract them to the
  area
• Permission marketing
• Design targeting issues
• Number of links, order of links, text or html
• Content targeting issues
• Content type (health, financial, etc.)

27
Data

• Three months of e-mails, 1048 users
• E-mail file e-mail date, number of links, order
  of links, link content, html or text
• User file when received, by whom (registration
  data), which links clicked (cookies)
• Sample 11,475 observations
• 7 response rate for links
• 36 click on more than one link

28
Models

• No heterogeneity
• Person heterogeneity
• Person, E-mail and Link heterogeneity (Full
  Model)

29
Predictive Ability
Predicted
Click
False Positives
Click
Actual
False Negatives
False Negative Fraction c/(cd), False Positive
Fraction b/(ab)
30
Predictive Ability Link Level ROC Curves
True Positive Fraction 1-FNF
False Positive Fraction
31
Predictive Ability Email Level ROC Curves
True Positive Fraction 1-FNF
False Positive Fraction
32
Results - Parameter Estimates Full Model

• Parameter Value Prob(m lt0)
• Design Variables
• Intercept (m0) -1.47 (1.0)
• Person Random Effects ( Std. l0i) 0.51
• E-mail Random Effects (Std. q0j) 0.45
• Link Random Effects (Std. g0k) 0.21
• E-mail Type (m1) 0.29 (0.48)
• Link Order (m2) -0.37 (1.0)
• Person Random Effects (Std. l2i) 0.49
• E-mail Random Effects (Std. q2j) 0.22
• Number of Links (m3) -0.02 (0.55)
• Person Random Effects (Std. l3i) 0.18

33
Parameter Estimates
Dirichlet Process Precision parameters User a1
103 gt 61 clusters Email a2 114 gt 65
clusters Links a3 383 gt 383 clusters
34
Link Level Predictions - Calibration Data
35
Link Level Predictions - Validation Data
36
E-mail Level Prediction - Calibration Data
37
E-mail Level Predictions - Validation Data
38
Optimization Model Overview

• Editorial content is fixed on a given day.
• n links available for k positions, n k
• How many links to include, what content to
  include, and how should it be ordered?
• Objective
• Maximize the expected number of click-backs to
  the site
• Maximize the likelihood of returning to the site

39
Optimization Procedures

• Alternative 1 Complete Enumeration
• With many links, computational constraints
• Alternative 2 Assignment Algorithm

40
Optimization Objective Function

• Maximize expected number of click-throughs to
  site
• Let xij 1 if link i is in position j
• Let pij be the probability of click through if
  link i is in position j
• Maximize Objective function
• Maximize likelihood of at least one click-through
• Minimize Objective function

41
Optimization Model

• Step 1
• Maximize Obj (x p(x)k)
• Subject to
• Assignment algorithm provides exact solution
• Step 2
• Maximize over k1, , n.

42
Heuristic Approaches

• Original - No change in content or order
• Greedy - No change in content, order highest
  utility first
• Order - No change in content, optimize order
• Optimal - Optimize content (number of links) and
  order (our procedure)

43
Optimization Results
44
Optimization Results

• Objective At Least One Click
• Optimal leads to 56 increase in at least one
  click.
• Re-ordering gives 52 improvement, content
  selection is the balance.
• Optimal improves over Order for 43 of e-mails
  (those adverse to clutter).
• Greedy and Order are similar, however for users
  who have high positive effect for order (scroll
  to bottom), Greedy does poorly (one user went
  from 81 to 43).
• Objective Expected Number of Clicks
• Similar results

45
Optimization Results
46
Conclusions

• Modeling link response
• Varies with content (information) and design (how
  much, what order)
• Heterogeneity in persons, links, and e-mails
• E-targeting
• Potential to considerable enhance clicks (and
  presumably advertising revenue and loyalty)
• Our approach can be applied to both internal and
  external targeting strategies
• Our approach can also be applied to e-tailing

47
Future

• Targeting
• Products and services for purchases
• Advertising
• E-grocers (features, displays, prices)
• How much is a feature worth?
• Other areas
• On-line choice processes
• Agent queries

48
Dirchlet Process Moments

• EG(B)EG0(B)
• and VarG(B)G0(B)(1-G0(B))/a

49
Full Conditionals for Individual Level Model
Normal Heterogeneity

• Standard Case (Simple Model)

50
Dirichlet Process Priors

• A c.d.f., G on Q follows a Dirichlet Process if
  for any measurable finite partition of (B1,B2,
  .., Bm), of Q, the joint distribution of the
  random variables
• ( G(B1), G(B2), , G(Bm)) is
• Dirichlet(aG0(B1), ., aG0(Bm)),
• where, G0 is a the base distribution and a is
  the precision parameter