MANAGEMENT SCIENCE informs ®

Vol. 52, No. 4, April 2006, pp. 597–612 doi 10.1287/mnsc.1060.0514

issn 0025-1909
eissn 1526-5501
06
5204
0597 © 2006 INFORMS

Machine Learning for Direct Marketing

Response Models: Bayesian Networks with
Evolutionary Programming
Geng Cui
Department of Marketing and International Business, Lingnan University, Tuen Mun, N.T., Hong Kong,
gcui@ln.edu.hk

Man Leung Wong

Department of Computing and Decision Sciences, Lingnan University, Tuen Mun, N.T., Hong Kong,
mlwong@ln.edu.hk

Hon-Kwong Lui
Department of Marketing and International Business, Lingnan University, Tuen Mun, N.T., Hong Kong,
hklui@ln.edu.hk

M achine learning methods are powerful tools for data mining with large noisy databases and give
researchers the opportunity to gain new insights into consumer behavior and to improve the performance
of marketing operations. To model consumer responses to direct marketing, this study proposes Bayesian net-
works learned by evolutionary programming. Using a large direct marketing data set, we tested the endogeneity
bias in the recency, frequency, monetary value (RFM) variables using the control function approach; compared
the results of Bayesian networks with those of neural networks, classification and regression tree (CART), and
latent class regression; and applied a tenfold cross-validation. The results suggest that Bayesian networks have
distinct advantages over the other methods in accuracy of prediction, transparency of procedures, interpretabil-
ity of results, and explanatory insight. Our findings lend strong support to Bayesian networks as a robust tool
for modeling consumer response and other marketing problems and for assisting management decision making.
Key words: direct marketing; Bayesian networks; evolutionary programming; machine learning; data mining
History: Accepted by Jagmohan S. Raju, marketing; received May 12, 2004. This paper was with the authors
6 months for 4 revisions.

1. Introduction (CART), and latent class regression, in a tenfold cross-

Machine learning is an innovative method that can
potentially improve forecasting models and assist
management decision making. Direct marketing,
which relies on building accurate predictive mod-
els from databases, is one of the areas that can
benefit from such applications. As more companies
adopt direct marketing as a distribution strategy,
spending in this channel has grown in recent years,
making consumer response modeling a top prior-
ity for direct marketers to increase sales, reduce
costs, and improve profitability. In addition to the
conventional statistical approach to forecasting con-
sumer purchases, researchers have recently applied
machine learning methods, which have several dis-
tinctive advantages for data mining with large noisy
databases. In this study, we adopt an innovative
machine learning method—Bayesian networks (BNs)
learned by evolutionary programming (EP)—to model
responses to direct marketing. We compare the results
of BNs with other benchmark methods, including
neural networks, classification and regression tree
597
validation with a large data set. The results suggest
that BNs have distinctive advantages, including accu-
rate prediction, transparent procedures, interpretable
results, and greater explanatory power.
1.1. The Statistical Methods
Because of budget constraints, most direct marketers
only contact a preset percentage (e.g., 20%) of the
names in a company’s database. Thus, the primary
objective of modeling consumer responses in direct
marketing is to identify customers who are most
likely to respond. Researchers have developed many
direct marketing response models using consumer
data. One of the classic models, known as the recency,
frequency, monetary value (RFM) model, determines
the likelihood of consumers responding to a direct
marketing promotion based on the recency of the
last purchase, the frequency of purchases over the
past years, and the monetary value of a customer’s
purchase history (Berger and Magliozzi 1992). Other
consumer demographic and psychographic variables,
 Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
598
credit histories, and purchase patterns may help build
to more sophisticated models that can improve the
understanding of consumer responses and the accu-
racy of purchase prediction.
Until recently, statistical methods such as logistic
regression and discriminant analysis have dominated
the modeling of consumer responses to direct mar-
keting (Berger and Magliozzi 1992). Although statisti-
cal methods can be very powerful, they make several
stringent assumptions on the types of data and their
distribution, and typically can only handle a limited
number of variables. Regression-based methods are
usually based on a fixed-form equation, and assume a
single best solution, which means that researchers can
compare only a few alternative solutions manually.
Further, when the models are applied to real data, the
key assumptions of the research methods are often
violated (Bhattacharyya 1999). Recently, researchers
have developed several more sophisticated models,
including beta-logistic models (Rao and Steckel 1995),
tree-generating techniques such as CART and CHAID
(Haughton and Oulabi 1997), and the hierarchical
Bayes model (Allenby et al. 1999). A number of stud-
ies have addressed the selection and endogeneity
biases in the existing models to improve predictive
accuracy (Bitran and Mondschein 1996, Gönül et al.
2000).
In recent years, the rapid accumulation of customer
and transactional data has resulted in very large
databases, and the voluminous amount of consumer
data provides unique opportunities for researchers
to use data-mining methods to gain insight into
consumer behavior. For instance, in addition to the
RFM variables, researchers have used consumer life-
time and transaction variables to improve the per-
formance of models (Bhattacharyya 1999, Venkatesan
and Kumar 2004). However, the increasing amount
and variety of customer data may render impossible a
manual solution to the optimization of response mod-
els (Bitran and Mondschein 1996). How to take advan-
tage of the different types and increasing volumes of
customer data to assist management decision making
presents new challenges. Innovative methods such as
machine learning allow researchers to perform data
mining with large databases to provide decision sup-
port to managers.
1.2. Machine Learning
Machine learning refers to computer-based methods
that can extract patterns or knowledge from data and
perform optimization tasks with minimum human
intervention. Most of these methods have their roots
in artificial intelligence and dynamic programming.
Machine learning methods have been adopted in
many fields as effective data-mining tools to discover
“interesting,” nonobvious patterns or knowledge hid-
den in a database that can improve the bottom
Management Science 52(4), pp. 597–612, © 2006 INFORMS
line. These methods include association rules, deci-
sion trees, neural networks, and genetic algorithms.
Business researchers have adopted some of these tech-
niques to solve classification problems, such as pre-
dicting bankruptcy and loan default and modeling
consumer choice (Hu et al. 1999, West et al. 1997).
Such methods can also be very useful in learning new
knowledge when researchers have observable data
but the model structure is unknown.
Artificial neural network (ANN), a procedure that
mimics the processes of the human brain, is among
several innovative methods that have been used
to model consumer responses to direct marketing
(Baesens et al. 2002, Zahavi and Levin 1997). In com-
parison with the statistical approach, simple forms
of ANNs are free from the assumptions of normal-
ity or complete data, and are thus particularly robust
for handling noisy data, including cases in which the
number and type of attributes vary (Michie et al.
1994). Moreover, neural networks are not subject to
the linearity assumption or the highly parametric
structure associated with models in a small-data set-
ting. They can explore complex structures to find
interactions, nonlinearities, and nonlinear interac-
tions, and are good at pattern discovery (Warner
1997). Given a sufficiently large amount of data, neu-
ral networks may offer better solutions to complex
optimization problems.
When Zahavi and Levin (1997) applied ANNs to
direct marketing, ANNs did not perform any bet-
ter than logistic regression. They attribute the prob-
lem of overfitting by back-propagation ANNs to the
complexity of the procedure, which typically starts
with selecting input and output variables, deter-
mining the number of hidden layers and hidden
nodes, and adjusting the weights of the nodes, all
by trial and error. In an attempt to solve this prob-
lem, several researchers have recently developed a
Bayesian approach to learning neural networks using
the Markov chain Monte Carlo (MCMC) method
(Neal 1996, Warner 1997). Using a set of hyperparam-
eters to select the appropriate priors and to repre-
sent the noise in the data, relatively complex struc-
tures are able to model the data while minimizing
overfitting (Warner 1997). Baesens et al. (2002) tested
the Bayesian approach to neural networks with direct
marketing data and produced positive results.
Despite these improvements and potential benefits,
machine learning still has limited applications in mar-
keting research, and is not short of skeptics. Such
methods have yet to make the necessary improve-
ments to allow marketing researchers to take advan-
tage of the features that they have to offer. First,
although methods such as ANNs can model complex
nonlinear structures, they lack a suitable topology
Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
Management Science 52(4), pp. 597–612, © 2006 INFORMS
of networks or method of knowledge representa-
tion to describe the relationships among the variables
(Zahavi and Levin 1997). Second, ANNs can generate
empirical results that are comparable to those of the
statistical approach, but its “black-box” procedures
learn complex relationships at an “unconscious” level
and are not transparent, thus making it difficult to
understand how a certain solution has been derived.
Third, their results are not easy to interpret to offer
managerial insight (Nakhaeizadeh and Taylor 1997,
West et al. 1997).
1.3. Objectives of the Study
In this study, we propose the innovative machine
learning method of BNs learned with EP to model
consumer responses to direct marketing. In this data-
mining task, we focus on learning a predictive model
by integrating the lifetime and consumer transaction
variables with the RFM variables to forecast con-
sumer purchases. For structural analysis, we use BNs
as a universal approximator to represent the model
structure, and then adopt EP as a stochastic search
algorithm to learn the optimal BN model. To eval-
uate the model fitness and minimize overfitting, we
apply the minimum description-length (MDL) met-
ric. In the following sections, we first elaborate the
advantages of BNs in representing model structures
and EP as an efficient optimization algorithm. Sec-
ond, we adopt the “control variable” approach to test
the endogeneity bias in the RFM variables. Third,
we test BNs learning with a large direct marketing
data set and compare the results with those of ANNs,
CART, and latent class regression in a tenfold cross-
validation. Fourth, we discuss the results and the
advantages of BNs in terms of the accuracy of clas-
sification, transparency of procedure, and interpreta-
tion of the results. Finally, we explore the managerial
599
take on any structure for a model. Such “freedom of
expression” allows the researcher to explore complex
relationships among the variables to discover new
knowledge, making BNs an ideal tool for data mining
(Heckerman 1997).
The main task of BNs is to decompose a joint prob-
ability distribution into a set of local distributions.
The network topology based on the independence
semantics specifies how to combine the local distri-
butions of the variables to obtain the joint proba-
bility through the nodes in the network (Haddawy
1999). The symmetric nature of conditional probabil-
ity allows researchers to perform prediction and diag-
nosis and to solve classification problems. In the past
decade, BNs have slowly made inroads into man-
agement research. For instance, marketing researchers
have adopted BNs to model strategic planning for
new products (Cooper 2000) and consumer com-
plaint behavior (Blodgett and Anderson 2000). EP,
an efficient stochastic optimization algorithm devel-
oped in the field of artificial intelligence, has been
adopted by researchers to identify optimal solutions,
but both BNs and EP are relatively new to manage-
ment researchers. It takes a great deal of theoretical
insight into these two methods to understand their
combined benefits.
2.1. Introduction to Bayesian Networks
By definition, a BN treats a research problem as mod-
eled by a list of variables and encodes the joint prob-
ability distribution of these variables:


P N1      Nn  = P Ni
Ni  (1)
i
First, a BN, B, has a qualitative part represented by a
directed acyclic graph (DAG) that depicts the condi-
tional independence among the variables in a domain
implications, potential applications of BNs in market- U = N1      Nn
and encodes the joint probability
ing research, and directions for further development. distribution (Pearl 1988). The network uses the vari-
ables as nodes to represent the relationships of depen-
2. Learning Bayesian Networks dency and independence among them. As is shown
Based on the well-developed Bayesian probability in Figure 1, each node in the graph corresponds to a
theory proposed about 250 years ago, BNs consti- variable in the domain. An edge Nj → Ni in the graph
tute a method of formal knowledge representation describes a parent and child relation in which Ni is
that has been around since the 1960s. In the last
two decades, the development of computers, algo- Figure 1 A Bayesian Network Model of Customer Complaints
rithms, and software has made it possible to exe-
P(rc) = 0.15 P(ui) = 0.01
cute realistic BN models (Jensen 1996, Pearl 1988).
Regular Unhappy
Since then, BNs have made significant strides in many customer (rc) incident (ui)
fields, such as software engineering, space naviga-
tion, and medical diagnosis (Haddawy 1999). Like P(sr|rc, ui) = 0.99
Repeat Service
ANNs, the BN approach is free from the assump- business (rb) recovery (sr)
P(sr|~rc, ui) = 0.90
P(sr|rc, ~ui) = 0.97
tions of data types and their normality and can effec-
P(rb|rc) = 0.6 P(sr|~rc, ~ui) = 0.03
tively handle nonlinearity, which is often associated
P(rb|~rc) = 0.05 Happy
with data mining using large databases. Moreover, customer (hc)
P(hc|sr) = 0.7
P(hc|~sr) = 0.01
BNs require no a priori model formulation and can
 Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
600 Management Science 52(4), pp. 597–612, © 2006 INFORMS

the child and Nj is the parent. An edge specifies a
dependency between Ni and Nj . All of the parents
of Ni constitute the parent set of Ni , which is denoted
by Ni . Bayes’ rule is used to update the conditional
probabilities given evidence.
Overall, BNs offer a formalism that can directly rep-
resent a complex distribution concisely and efficiently.
given each possible combination of states of its par-
ents. If a node contains no parent, then the marginal
probabilities of the node are used (Pearl 1988). The
example in Figure 1 suggests that the probability
of a happy customer (hc) is the joint probability of
the local probabilities of other events, such as ser-
vice recovery (sr), an unhappy incident (ui), a repeat
Let the domain variables be U = N1      Nn
, the joint customer (rc), and repeat business (rb). The probabil-
probability table grows exponentially with the num-
ber of variables, and U does not need to be very
large before the table becomes intractably large. A BN
over U is a compact representation of the joint prob-
ability distribution, and its information can be calcu-
lated using Equation (1). Figure 1 is a hypothetical
BN model for handling customer complaints with five
variables: regular customer, unhappy incident, ser-
vice recovery, repeat business, and happy customer.
Because all the variables are binary, the joint probabil-
ity distribution table should have 25 − 1 = 31 entries.
However, the BN has only 10 probability values, and
thus 21 values are saved. If there are more domain
variables, the amount of values saved will be much
larger.
2.1.1. Conditional Dependence. Secondly, based
on a model structure, the quantitative part of BNs esti-
mates the conditional probabilities for the variables
in the model. BNs operate on the assumption of con-
ditional independence. Let U be the set of variables
in the domain and P be the joint probability dis-
tribution of U . Following Pearl’s notation, a condi-
tional independence relation is denoted by IX Z Y ,
where X, Y , and Z are disjoint subsets of the variables
in U . This notation says that X and Y are condition-
ally independent given the conditioning set Z. Thus,
in a three-node network, one of the variables acts as
a “virtual control” for the relationship between the
other two. Formally, a conditional independence rela-
tion is defined as in Pearl (1988) as
P x
y z = P x
z where P y z > 0
where x, y, and z are any value assignments to the
set of variables X, Y , and Z, respectively. A condi-
tional independence relation is characterized by its
order, which is simply the number of variables in
the conditioning set Z. This is also referred to as the
d-separation, which is used to determine the condi-
tional dependence and independence among the vari-
ables, and is conditional on the state of the other
variables (Pearl 2000). Probability calculus is used to
quantify the relationships of dependence represented
ities for all the nodes in the Bayesian network can be
calculated. For example, as P rc = 015 and P ui =
001, then
P rb = P rb
rcP rc + P rb
∼rcP ∼rc
= 06 × 015 + 005 × 1 − 015 = 01325 (3)
The probabilities for the other nodes in the network
can be derived in a similar fashion.
2.1.2. Inference and Causality. With a BN model
at hand, probabilistic inference can be performed to
predict the outcome of certain variables based on
the observation of others. Intuitively, an edge in a
BN expresses the notion of “interaction” between
variables, and a missing edge represents a missing
relation between two variables. Through conditional
dependencies and in probabilistic terms, the DAG
gives a lucid representation of the dependencies and
irrelevancies among the variables embedded in a net-
work (Chiogna 1997). For example, the service recov-
ery (sr) event may lead to the happy customer (hc)
event, and the strength of this relationship would be
represented by the conditional probability P hc
sr.
The posterior probability of observing the event of
service recovery when we see a happy customer can
be obtained using the well-known Bayes’ rule
P hc
srP sr
P sr
hc = = 09383 (4)
P hc
Although the formal definition of a BN is based on
conditional independence, in practice it is often con-
(2) structed using the notions of cause and effect, which
makes it a powerful tool for the identification and
analysis of the structural relationships among vari-
ables (Heckerman 1997). With data on intervention
or similar knowledge, researchers can explicate the
causal relationships among variables (Pearl 2000). An
edge, Nj → Ni , may represent causality, with Nj being
the cause and Ni being the effect. Their conditional
dependencies can help distinguish causation from
mere correlation or association, and can lead to the
inference of causality on a solid mathematical basis.
Consider the BN shown in Figure 1, in which the par-
by a BN.
In a BN, each node has a conditional probability
. Happy customer is independent of
distribution in the form of P Ni
Ni , which speci-
fies the probability of each possible state of the node
ent set of the node service recovery is regular customer
unhappy incident
regular customer and unhappy incident given service
recovery, whereas the effect of regular customer and
Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
Management Science 52(4), pp. 597–612, © 2006 INFORMS
unhappy incident is mediated by service recovery.
A model with such modularity helps to explain how
these probabilities change as a result of external inter-
vention. This is particularly useful for research on the
effect of management decisions or business strategies.
2.2. Learning Bayesian Networks Using
Evolutionary Programming
A BN can be constructed based on previous research,
as in Blodgett and Anderson (2000), or by eliciting
knowledge from domain experts, as in Cooper (2000).
However, building BNs based on previous informa-
tion or expert knowledge can be difficult, because
such information may not be available. To reduce the
imprecision due to subjective judgments, researchers
can learn a BN from collected data, instead of fitting a
specified model to the data. For BNs to “learn” from
the observed data a probability distribution that can
best describe the relationship among the variables,
researchers have devised various learning algorithms,
such as the genetic algorithm (Larrañaga et al. 1996).
In this study, we propose to learn BNs using EP. Fig-
ure 2 provides a graphic representation of the process
in which EP learns BNs to solve prediction and clas-
sification problems.
First, data reduction is undertaken to extract the rel-
evant variables from the database. Second, researchers
need a fitness measure to assess the goodness of BN
models (Figure 2). Such measures may be derived
from the Bayesian information criterion (BIC) or
Figure 2 Data Mining Using Bayesian Networks Learned by Evolutionary Programming
Learning Bayesian network structure and conditional probabilities
Generate an initial
Data population, P, of
preparation DAGs. Evaluate
them using MDL
Max. no
of generations
yes
reached?
no
Generate new
DAGs from P and
evaluate them
Select DAGs from
P and new DAGs
Store the selected
DAGs in P
601
Akaike information criterion (AIC). In this study, we
employ the MDL metric (Lam and Bacchus 1994),
which is discussed in §2.2.1. With the defined metric,
learning BNs is formulated as a search problem, and
here we use EP to tackle the problem. We describe EP
in §2.2.2 and discuss our learning algorithm in §2.2.3.
Section 2.3 explains how the learned BN is used to
perform prediction and classification.
2.2.1. The Minimum Description-Length Metric.
An appropriate fitness measure is critical for model
evaluation and selection. Despite their advantages in
handling nonlinearity and learning complex models,
BNs based on a uniform prior may overfit a partic-
ular data set (Lam and Bacchus 1994). In the worst
case, the computation of the posterior probabilities in
complex network models becomes intractable. Com-
plex BNs require more probabilities and tremendous
computer space to store them, and suffer from the
conceptual disadvantage that a model with complex
structures makes it difficult to understand and explain
the underlying relationships. As a result, simpler
models are preferred if they are sufficiently accurate.
Rissanen (1978) proposed the MDL principle, and
Lam and Bacchus (1994) subsequently adopted it for
evaluating BNs. This metric is rooted in information
theory, and is equivalent to the Bayesian scoring func-
tion or the BIC (Hansen and Yu 2001). In other words,
when the number of samples increases, the learned
model converges to the underlying true distribution
Calculate the Classification
Select the best conditional and
DAG from P probabilities in prediction
the best DAG
Bayesian
network
output input
 Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
602
of the data with a probability equal to one. As
the MDL imposes a penalty on model complexity, the
burden of proof falls on complex models. Thus, the
MDL metric strikes a balance between model accu-
racy and simplicity, and has been employed as the
fitness function to evaluate BNs. It effectively serves
as a mechanism to control overfitting (Hansen and Yu
2001).
Motivated by information coding (Rissanen 1978),
the MDL principle assumes that a collection C of data
items is given, and that it is necessary to place this
collection in computer storage. The encoded collection
is referred to as the total description length, which is
defined as the sum of the length of the compressed
version of data C and the description length of the
model. The MDL metric measures the total descrip-
tion length Dt (B) of a BN structure B. The metric dic-
tates that the optimal model to explain a collection of
data minimizes the total description length. Let N =
Management Science 52(4), pp. 597–612, © 2006 INFORMS
the accuracy of a network. The compressed version
of the data includes the values of x1  x2      xn and
the errors e1  e2      en . As the storage size that is
required for x1  x2      xn is fixed for all the models, if
one model has a shorter data description length than
another model, then the storage size for the errors
of the first model is smaller than that of the second
model. Thus, a model is more accurate if the corre-
sponding data description length is smaller.
To encode a BN, we need to encode the network
topology of the graph (the list of parents for each
node) and the set of conditional probabilities associ-
ated with each node. For a BN with n nodes, it is suf-
ficient to encode a list of the parents of each node and
a set of conditional probabilities for each node. For a
node with k parents, we need k log n bits to encode
the list of its parents. The encoding for the conditional
probabilities depends on the number of parents and
the number of values that the variables take.
N1      Nn
denote the set of nodes in a BN and Ni
denote the set of parents of node Ni . The total descrip-
tion length of a BN Dt  is the sum of the description
lengths of each node
 advantage of the MDL metric is that it balances accu-
Dt B = Dt Ni  Ni  (5)
Ni ∈N
The total description length Dt  is based on two
components: the network description length Dn  and
the data description length Dd . Thus, the MDL score
of a model depends on the sample size of the data
and the complexity of the model
Dt Ni  Ni  = Dn Ni  Ni  + Dd Ni  Ni 
The formula for the network description length is
as follows.

other criteria such as AIC and BIC, MDL is a robust
Dn Ni  Ni  = ki log2 n + dsi − 1 sj  (7)
j∈ Ni
where ki is the number of parents of variable Ni , Si is
the number of values Ni , Sj is the number of values
of a particular variable in Ni , and d is the number
of bits required to store a numerical value. This is the
description length for encoding the network structure.
The first part in the addition is the length for encod-
ing the parents, and the second part is the length
for encoding the probabilities. The model description
length measures the simplicity of the model.
The formula for the data description length Dd  is
 M Ni 
Dd Ni  Ni  = MNi  Ni  log2
Ni ∈ Ni
MNi  Ni 
where M· is the number of cases that match a partic-
ular instantiation in the database. This is the descrip-
tion length for encoding the data, which measures
As BNs encode the data as a joint distribution of
probabilities, which can be quite complex, the MDL
is particularly suitable as a method to select BN mod-
els in data mining with large databases. The key
racy and simplicity of models. A simpler network is
preferred, providing it is sufficiently accurate, but if
there is no simpler network that is accurate enough,
then the metric allows a more complex network to be
induced. The MDL metric thus guides the search for
a more accurate network by increasing its topologi-
cal complexity (Lam and Bacchus 1994). On the other
hand, it may select a simpler network with fewer
(6) errors over a more complex model that perfectly fits
the data. Viewed in this light, the MDL provides an
effective mechanism to minimize the overfitting of the
data. Recent experiments suggest that compared with
metric and is increasingly used for model selection
(Hansen and Yu 2001, Mitchell 1997).
2.2.2. Evolutionary Programming. Evolutionary
computation refers to a family of computational meth-
ods that perform machine learning and function opti-
mization by simulating the natural evolution process
based on the Darwinian principle. Genetic algorithms
(GAs), genetic programming (GP), evolution strat-
egy (ES), and evolutionary programming (EP) are all
examples of evolutionary computation methods. They
are robust and parallel search tools for identifying
accurate models from all possible hypotheses and dis-
covering new knowledge from noisy data. Their main
differences are in the evolution model assumed, the
 (8)
evolutionary operators employed, and the selection
method or the fitness functions used.
This study adopts EP because of its several distinct
advantages. First, whereas GAs focus on binary bits
and GP uses tree structures, EP does not require any
Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
Management Science 52(4), pp. 597–612, © 2006 INFORMS 603

type of model structure. The model can be a binary Table 2 The Algorithm for Evolutionary Programming to Learn
string, a tree structure, or any other shape, and can Bayesian Networks
evolve into new structures during the evolutionary • Set t to 0.
process. Second, unlike GAs that use reproduction, • Set I to 0.
crossover, and mutation operators, mutation is the • While I is smaller than the population size P S,
only genetic operator for evolution in EP. Whereas
Let N be the set of all nodes.

Let B be a Bayesian networks without any edges.
crossover operation may lead to invalid models such
For each Ni in N,
as recursive models, mutation operator alone in EP
Randomly generate an integer k from 0 to 5.
allows the simultaneous modification of all the nodes
(variables), which by itself is powerful and makes
crossover redundant. Third, instead of emulating the
specific genetic operators observed in nature, muta-
Insert B into the initial population Popt.
tions in EP preserve the behavioral similarity of
Increase I by 1.
the parents and their offspring models (Fogel 1994).
A “child” model is generally similar in behavior to
the “parent” model, with slight variations. Thus, EP
represents a model of evolution at a higher level of
Each BN in Popt produces one offspring by performing a number
abstraction (Wong et al. 1999). Compared with the
GA methods to learn BNs (Larrañaga et al. 1996), EP
is more flexible and explores wider search spaces to
The BNs in Popt and all new offspring are stored in the
. The size of Popt
 is 2 ∗ P S.
compare alternative models. Because of these unique
Conduct a number of pairwise competitions over all BNs in Popt
.
advantages, EP is an ideal search mechanism for opti-
mization purposes, and has produced more satisfac-
 with equal probability. Let Bij ,
tory results than GAs in both model performance and
computing efficiency when applied to learning BNs
Select P S BNs with the highest scores from Popt
 and store them
(Fogel 1994, Wong et al. 1999).
A typical process of EP is outlined in Table 1. A
Increase t by 1.
set of models is randomly created to make up the
initial population. Each model is evaluated by the fit-
ness function, and then each model produces a child
by mutation. There is a certain distribution of the
different types of mutation, which range from minor
to extreme, and minor modifications in the behav-
ior of offspring occur more frequently than substan-
tial modifications. The offspring are also evaluated by
the fitness function, and then tournaments are per-
formed to select the models for the next generation.
For each model, a number of rivals are randomly
selected among the parents and their offspring. The
tournament score of a model is the number of its
rivals with worse fitness scores than itself. The mod-
els with higher tournament scores are selected as the
Table 1 The Algorithm of Evolutionary Programming
Initialize the generation, t, to be 0.
Initialize a population of individual, Popt
Evaluate the fitness of all individual in Popt
While the termination criteria is not satisfied
Produce one or more offspring from each individual by mutation
Evaluate the fitness of each offspring
Perform a tournament for each individual
Put the individuals with high tournament scores into Popt + 1
Increase the generation t by 1
Return the individual with the highest fitness value
Note. Unlike statistical procedures, machine learning algorithms are mostly
written in logical language, rather than mathematical formulas.

Randomly select k nodes from N\Ni  without replacement.

For each selected node Nj ,
• If cycles are not generated by inserting the edge Ni ← Nj
into B, then add the edge Ni ← Nj into B.
• Each BN in the population Popt is evaluated with the fitness function
defined in Equation (5).
• While t is smaller than the maximum number of generation G and there
are opportunities for improvement,
of mutation operations. If the offspring has cycles, then delete the
edges of the offspring that invalidate the acyclic condition.
intermediate population Popt
Let Bi be the BN that is conditioned upon, and q opponents are
selected randomly from Popt
1 ≤ j ≤ q, be the randomly selected opponent BNs. The Bi gets one
more score if Dt Bi  ≤ Dt Bij , 1 ≤ j ≤ q.
in the new population Popt + 1.
• Let the BN with the lowest fitness value found in any generation of a
run be Bbest .
• Calculate the parameters of Bbest by using Equation (10).
• Return Bbest as the result of the algorithm.
population of the next generation. The population size
(the number of competing models) does not need to
be held constant. The process is iterated until the ter-
mination criterion is satisfied.
2.2.3. The Learning Algorithm. Our EP algorithm
for learning BNs is depicted in Table 2. Each individ-
ual represents a BN model, which is a DAG. First, a
set of BNs is randomly generated to make up the ini-
tial population. Each graph is evaluated by the MDL
metric described above. Then, each BN produces off-
spring by performing a number of mutations. The
probabilities of using one, two, three, four, five, or
six mutations are set to 0.2, 0.2, 0.2, 0.2, 0.1, and
0.1, respectively. The mutation operators modify the
edges of a DAG. If a cyclic graph is formed after the
mutation, edges in the cycles are removed to keep it
acyclic.
After generating the offspring models, they are
also evaluated by the MDL metric. The next gener-
ation of the population is selected among the par-
ents and offspring by tournaments. Each DAG B is
compared with other randomly selected DAGs, and

604
a tournament score of B equals the number of rivals
that B can beat, that is, the number of DAGs among
those selected that have higher MDL scores than B.
In our setting, q = 5. One half of the DAGs with the
highest tournament scores are retained for the next
generation. As is depicted in Table 2, the process is
repeated until the maximum number of generations
is reached, which depends on the complexity of the
network structure. If one expects a simple network,
the maximum number of generations can be set to a
lower value. The network with the lowest MDL score
emerges as the final solution.
Once the learning algorithm returns the BN model
with the lowest MDL score, the conditional probabil-
ities of the nodes in the network can be calculated as
follows.
Nijk + 1
P Ni = vik
Ni = wij  =  (9)
Nij + ri
where vik is a value of variable Ni , wij is an instantia-
tion of the parent set Ni , ri is the number of different
values of variable Ni , Nijk is the number of cases in
the database in which variable Ni has the value vik ,
and Ni is instantiated as wij with
ri
 for unobserved heterogeneity among subjects using
Nij = Nijk 
k=1
2.3. Bayesian Networks for Classification and
Prediction
As is depicted in Figure 2, learning BNs for classi-
fication takes two steps. First, for BN learning, the
EP algorithm automatically finds the directed edges
between the nodes to identify a network model that
can best describe the relationships based on the MDL
metric. Once the best network structure has been
identified, the conditional probabilities are calculated
based on the data to describe the relationships among
the variables. Second, the learned BN is used to gen-
erate a probability score for each example case or
unseen data for the purpose of cross-validation and
forecasting (see §§2.1, 2.2.3, and Equation (10)). The
probability score for the dependent variable (e.g., pur-
chase), which ranges from 0 to 1, is used for predictive
modeling. With these probability scores, researchers
can use a cutoff point (e.g., 0.5) to evaluate the error
rate of a predictive model. The predicted class or
membership is then compared with the actual data
to evaluate the predictive accuracy of the model and
to make sales forecasts by validating the results on a
testing data set. Alternatively, researchers can exam-
ine the percentage of true positives in the top deciles
of the testing (validation) data to arrive at the predic-
tive lift of the model.
Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
Management Science 52(4), pp. 597–612, © 2006 INFORMS
3. Results
3.1. The Data Sets and Experiments
To test the proposed methods for modeling consumer
responses to direct marketing, we perform experi-
ments to learn BNs with a direct marketing data set.
The Direct Marketing Education Foundation provided
the data from a U.S.-based catalog direct marketing
company that sells multiple product lines of general
merchandise that range from gifts and apparel to con-
sumer electronics. The company sends regular mail-
ings to its list of customers, and this particular data
set contains the records of 106,284 consumers. Each
customer record contains 361 variables, including
purchase data from recent promotions and the cus-
tomer’s purchase history over a 12-year period. In
a recent promotion, every customer in this data set
received a catalog. This promotion achieved a 5.4%
response rate, which represents 5,740 customers who
made purchases from this catalog.
In this study, we compare BN’s performance with
that of several other methods that are known for their
ability to solve classification problems. They include
two other machine learning methods, ANN using
Bayesian learning and MCMC (Neal 1996), and the
classification tree by CART (Haughton and Oulabi
1997). Recently, latent class models, which control
(10)
latent or unknown groups, have become increasingly
popular in marketing research, and thus serve as
another method for comparison with BNs (Jain et al.
1990). We examine the performance of these classi-
fication methods by testing the predictive ability of
the models learned from the training data on unseen
cases, that is, the testing (validation) data.
In direct marketing applications, simple error rate
may not be the most appropriate method for assess-
ing classifier performance. First, despite the large size
of the data set, the percentage of true positives is very
small (5.4% in this case). Second, simple error rate
assumes no variance in the cost of different misclas-
sification errors (Baesens et al. 2002). For direct mar-
keting models, false negatives are much more costly
than false positives. The loss from false positives is the
cost of mailing, but the opportunity cost from false
negatives—the loss of potential sales (US$80 on aver-
age in this case) and profit—is often much greater.
Furthermore, due to budget constraints, typically only
the names in the top two deciles or the 80th percentile
(those with the highest probability to respond) will
receive a catalog (Berger and Magliozzi 1992). Thus,
the cumulative response lift in the top two deciles of
the file (testing data sets) is used to compare the per-
formance of these methods. It is the ratio ∗ 100 of the
number of true positives (TPs) in a decile identified by
the proposed model versus the number of TPs identi-
fied by a random model, which is the number of TPs
Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
Management Science 52(4), pp. 597–612, © 2006 INFORMS
divided by the number of deciles (10). For instance, a
model with a top decile lift of 200 is said to perform
twice as well as a random model.
3.2. Dealing with Endogeneity
In this data-mining task, we are interested in explor-
ing the insight that can be gained by integrating life-
time and transaction variables with RFM variables.
First, nine variables are selected by logistic regres-
sion using the forward selection criterion p = 005:
recency (Recency), which is the number of months
that have elapsed since the last purchase; the fre-
quency of purchase in the last 36 months (Frequency);
the monetary value of purchases in the last 36 months
(Monetary value); the average order size (Ordsize);
lifetime orders (number of orders placed, Liford); life-
time contacts (number of mailings sent, Lifcont); and
whether a customer typically places order by tele-
phone (Tele); makes cash payment (Cash); or uses
the “house” credit card (Hcrd) issued by the catalog
company. Although every customer in this data set
received a catalog from the company in the current
mailing, the data do not have information on prior
selection by the management for each of the previous
mailings. We cannot explicitly model management
selection with this data set, but the model includes the
number of lifetime contacts (Lifcont), which indicates
the previous selection by the management.
Another critical issue is the possible endogeneity
of the RFM variables. In a structural equation model
y = )x + *, a predictor variable must be uncorrelated
with the error of the model to infer the causality of x
on y. If x is correlated with the model error *, this
variable is said to be endogenous, and its parameter
estimates may be biased. Such problems are due to
the omitted variables embedded in the error, which
simultaneously affects both the y and x variables.
Direct marketers often use RFM variables to predict
future purchases, yet RFM variables are based on pre-
vious responses from these households. In this sense,
the RFM variables may be endogenous and their
parameter estimates biased due to the correlations
between RFM variables and the error of the model.
This is a common problem that is often associated
with reoccurring data and arises from the lack of
empirical data to control for endogeneity.
Unlike regression methods, which focus on the con-
ditional probability, BNs do not specify a model struc-
ture, but learn a joint probability distribution among
all the variables simultaneously from the observed
data. Thus, BNs do not suffer from the endogene-
ity bias. Both ANN and CART estimate a condi-
tional distribution and may suffer from a potential
605
endogeneity bias. However, no solution to this prob-
lem has been devised for these two methods. Readers
are advised to proceed with caution in interpreting
their results. As for latent class regression, which
assumes a logit model, endogeneity correction is nec-
essary to remove such bias and to produce consis-
tent parameter estimates. In the existing literature,
researchers have adopted several solutions to this
problem, including the instrumental variable method
(Gönül et al. 2000) and the control function approach
(Blundell and Powell 2004). As our dependent vari-
able is binary, we adopt the control function approach
to test for endogeneity bias, which is accomplished by
adding the residuals of the endogenous variables into
the model as control variables.
Following the procedures in Blundell and Pow-
ell (2004), we first run a parametric reduced-form
regression to compute the estimates of endogenous
RFM variables on the whole data set. In the second
stage, the residuals of the reduced-form regressors are
included as covariates in the binary response model
to account for their endogeneity. In Table 3, the first
three columns are the reduced-form estimates for the
RFM variables with two lifetime variables as covari-
ates. Given their adjusted R-squares, the explanatory
power of the three reduced-form equations is fairly
high. The fourth column refers to the model without
any adjustment for endogeneity. Except for monetary
value, recency and frequency have very small coeffi-
cient estimates, although they are statistically signif-
icant. Despite the large data set N = 106280, the
overall fit of the uncorrected model is statistically
insignificant p = 0620, indicating a potential endo-
geneity bias in the RFM variables. The last column
includes the residuals of the RFM variables as the con-
trol variables. The coefficient estimates change drasti-
cally, and thus the effect of correcting for endogeneity
is obvious given the improvement in model fitness
p = 0021. For the endogeneity tests, we employ the
asymptotic t-test developed by Smith and Blundell
(1986). The significant results of the t-tests reject the
null hypothesis of exogeneity for the RFM variables.
To assess the effect of endogeneity correction on
predictive performance, we test the corrected RFM
model by including the residuals of the RFM vari-
ables in the latent class logit model, perform a tenfold
cross-validation, and compare the results with those
of the uncorrected model. The use of the control func-
tion approach in the model produces a top decile lift
of 401, which is a significant improvement over the
uncorrected model (with a top decile lift of 334) and
the model corrected with the instrumental variables
(with a top decile lift of 397). Thus, the control func-
tion approach is adopted for subsequent experiments
with the latent class method.
 Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
606 Management Science 52(4), pp. 597–612, © 2006 INFORMS

Table 3 Results of the Endogeneity Tests

Reduced-form regression Logit model

Models and tests/variables Recency Frequency Monetary value Uncorrected Corrected

Constant 32650∗∗ 0969∗∗ 3427∗∗ −4068∗∗ −11225∗∗

0206 0022 0006 0097 1177
Recency — — — −0026∗∗ −0219∗∗
0002 0019
Frequency −6428∗∗ — 0379∗∗ 0025∗∗ −1297∗∗
0028 0002 0010 0056
Monetary value — — — 0368∗∗ 3511∗∗
0022 0217
Lifetime contacts −2353∗∗ −0193∗∗ 00545∗∗ — —
0118 0013 0001
Lifetime contacts2 0302∗∗ 0048∗∗ — — —
0015 0002
Lifetime orders 0668∗∗ 0121∗∗ −00085∗∗ — —
0011 0001 0000
Lifetime orders2 −0006∗∗ −0001∗∗ — — —
0000 0000
Adjusted R2 0345 0255 0464
Endogeneity test: Recency 9428∗∗ t 
Endogeneity test: Frequency −14486∗∗ t 
Endogeneity test: Monetary value −15110∗∗ t 
p = 0620 p = 0021

Notes. (1) The standard errors are shown in parentheses. (2) The endogeneity test is the asymptotic t-test.
∗∗
= significant at the 0.001 level.

3.3. Results of Bayesian Networks transparency, and the MDL metric is effective in opti-
To learn BNs, discretization is performed by the recur- mizing BNs.
sive minimal entropy partitioning method for several We now examine the empirical results of BNs
continuous variables with a large number of values, to assess their ease of interpretation and manage-
including frequency and monetary value, to reduce rial insight. The DAG for the optimal BN model
the size of the data matrix. In the EP learning process, delineates the qualitative relationships among the
the number of BN models in each generation is set variables (Figure 3). Close examination of the DAG
at 100, with 50 parent models and 50 offspring mod- model reveals several structures. The three variables
els. The maximum number of generations is set at of recency, frequency, and monetary value form a net-
5,000, which is sufficient to allow even complex mod- work that has a direct effect on consumer responses to
els to converge. In a holdout validation experiment in the promotion, which confirms the explanatory power
which 90% of the data is allocated for training and of the RFM model, despite its simplicity. The two
the remaining 10% for validation, the BN model has
a top decile lift of 413, which is 4.13 times as good as Figure 3 A Directed Acyclic Graph Model for the Catalog Promotion
a random model.
The MDL scores of the models in the first gen- Cash Recency
eration are around 1,130,000. Then the MDL metric
Tele
of BNs declines as EP learns better network models.
The MDL metrics of subsequent models show signifi- Hcrd
cant improvement until the 870th generation, beyond Frequency Purchase
which the rate of improvement starts to level off. The
Lifcont
optimal BN model appears at the 1,045th generation,
with a MDL score of 1,047,550. As each generation Liford
compares 100 models (50 parent models and 50 off-
Monetary
spring models), the optimal solution emerges after the Ordsize
value
comparison of more than 50,000 models. These results Note. Due to limited space, the posterior probabilities and examples of con-
suggest that the performance of BNs is easily tractable ditional probabilities of this model are attached in Table 4 and Appendix A,
in the EP process and demonstrates a high degree of respectively.
Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
Management Science 52(4), pp. 597–612, © 2006 INFORMS 607

lifetime variables of lifetime contacts (Lifcont) and Table A3, at a high frequency level (4) consumers tend
lifetime orders (Liford), together with the RFM vari- to place telephone orders with a probability of 0.447,
ables, form a larger network. In addition, the trans- and at a lower frequency level (2) they are more likely
action variables of telephone (Tele), using the house to pay with the house card P = 0351. At a higher
credit card (Hcrd), and cash payment (Cash), together frequency level (4) they may place telephone orders
with recency and frequency, also help to explain and use the house card for payment at the same time
consumer responses. Overall, the RFM, lifetime, and P = 0304. These conditional probabilities help us
transactional variables together form a BN model to to understand the general relationships among the
predict consumer responses. variables.
Based on the DAG model in Figure 3, the poste- Furthermore, we can combine the posterior and
rior probabilities are generated for all nine variables conditional probabilities to perform inferences in a
(Table 4). These results define the quantitative rela- backward fashion to determine the combined effect
tionships among the variables, and help us to under- of several variables on purchase probability (Blod-
stand how they relate to consumer responses. First, gett and Anderson 2000). In other words, we can
the overall probability of response is 5.4%. The prob- reverse the question—“Given a customer’s purchase
ability of purchase is the highest when the value of from the promotion, what is the probability of a par-
recency is 2, but is also relatively high when the value ticular variable or value being the cause?” As for
of recency is 4 or 6, which means that recency has a the effects of lifetime contacts (Table A4), at low fre-
nonlinear effect on purchase. Higher values for fre- quency levels (2) lifetime contacts make little differ-
quency and monetary value lead to a higher purchase ence in consumer purchase, but at high frequency lev-
probability. This is also true for lifetime contacts, life- els (5) lifetime contacts increase the purchase proba-
time orders, and order size. The probabilities of using bility. These results can help identify the most attrac-
the house credit card, cash payment, and telephone tive customer groups and build consumer profiles
order are relatively low. Based on these results, we can for segmentation purposes. Based on the conditional
draw forward probabilistic inferences about the effect
probabilities, researchers can also suggest specific
of one variable on the likelihood of purchase given the
actions to improve the response rate of direct mar-
information of the other variables. For instance, given
keting campaigns. For instance, in Table A5, at a low
everything else we know about these customers, and
frequency level (1) the purchase probability is very
when the value of recency is 7 (the largest number of
low (0.018), and neither telephone orders nor pay-
months that have elapsed since the last purchase), the
ment with the house credit card makes a difference.
posterior probability of purchase is 0.018.
At the highest frequency level (5) the probabilities of
Moreover, associated with the graphic model are
placing a telephone order and using the house credit
the conditional probabilities for each of the variables
card increase substantially. These results suggest that
(Appendix A). According to Table A1, given the val-
ues of recency (2) and frequency (4), when monetary by integrating customer lifetime and transaction vari-
value is the highest (6) the conditional probability ables with the RFM variables, we can gain additional
of purchase equals 0.421, which is more than twice insight into consumer responses.
as high as when monetary value is lower (2 and 3).
3.4. Results of the Other Methods
Table A2 suggests that if consumers tend to use the
We adopt the Bayesian approach to learning neural
house card for payment, they will also use the tele-
network models and test different ANNs by adjust-
phone to place orders P = 0635. As is shown in
ing the number of hidden layers and hidden nodes
(Warner 1997). After repeated trials, an ANN model
Table 4 Posterior Probabilities of Variables in the Bayesian Network
Model
with one hidden layer and five hidden nodes (neu-
rons) returns the best model with the lowest error
Variables/values x =0 x =1 x =2 x =3 x =4 x =5 x =6 x =7 rate. Thus, the learning process is neither automatic
P (Purchase) 0.054 nor transparent. The neural network model depicts
P (Recency) 0.056 0.168 0.032 0.066 0.023 0.054 0.018 the unidirectional effects from the input variables to
P (Frequency) 0.018 0.045 0.063 0.092 0.137
P (Monetary value) 0.019 0.046 0.062 0.085 0.115 0.183
the hidden nodes and purchase. In terms of predic-
P (Lifcont) 0.036 0.050 0.048 0.058 0.085 tive accuracy, the ANN model has a top decile lift
P (Liford) 0.038 0.075 0.087 0.153 of 376 on the holdout validation data set. In addition,
P (Ordsize) 0.042 0.051 0.057 0.066 0.071 0.086
the model offers a set of weights or parameters for
P (Hcrd) 0.053 0.055
P (Cash) 0.057 0.040 the edges from the input variables to the five hid-
P (Tele) 0.050 0.070 den nodes 5 ∗ 9 = 45, the weights from the hidden
Notes. (1) Hcrd, Cash, and Tele are binary variables, and thus only two val-
nodes to the output, and the weights from bias (inter-
ues are recorded. (2) Due to the symmetric nature of the probabilities, the cept) to the hidden nodes and the output. The hidden
probabilities for nonpurchases are omitted. nodes can be viewed as latent factors or segments,
 Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
608 Management Science 52(4), pp. 597–612, © 2006 INFORMS

and the activation values (weights) of the variables Table 5 Coefficient Estimates by Latent Class Regression
vary across them. We may conclude that the hidden Variables Overall Wald p-value Class 1 Class 2 Wald(=) p-value
neurons help to explain consumer response. Although
Recency 0848 13301 0.000 0173 3136 32402 0.000
ANNs achieve a reasonable level of predictive accu- Frequency −0648 23883 0.000 0323 −3943 36681 0.000
racy, its empirical results do not formalize the rela- Monetary value 0001 0994 0.320 0001 −0001 11268 0.004
tionships among the variables in a user-friendly and Lifcont 1410 24292 0.000 0468 4606 170049 0.000
Liford 0687 89470 0.000 0140 2541 114781 0.000
comprehensible way, nor does it provide the oppor- Ordsize 0052 0762 0.380 0017 0170 2296 0.320
tunity to gain fresh insight into the problem. Even Hcrd −0243 0154 0.700 −0225 −0303 78662 0.000
for competent users of ANNs, its lack of transparency Cash −0052 2399 0.120 −0092 0086 9939 0.007
Tele −0124 1320 0.250 −0040 −0411 6890 0.032
and explanatory capability has been a major draw- Res. R −0902 14768 0.000 −0187 −3327 36560 0.000
back (West et al. 1997). Res. F 6673 26096 0.000 0940 26128 44249 0.000
CART is a recursive partitioning method that gen- Res. M 0007 0468 0.490 −0053 0208 1765 0.410
erates a tree model by “splitting the tree” at each R-square 0225 0048 0462
node. It uses the GINI index to determine how well
the splitting rule separates the classes contained in parameter estimates from Class 2, albeit a different
the parent node. Once the best split is found, CART profile. A researcher may suggest that heterogene-
repeats the search process for another child node, and ity across these two clusters may improve predictive
continues recursively until further splitting is impos- accuracy, but again the meaning of such unobserved
sible. Instead of deciding whether a given node is clusters and their effect on consumer response is not
terminal, CART grows the tree to the maximum size apparent to the average analyst and must be rational-
and then starts “pruning” it to examine smaller trees. ized by the researcher.
Finally, CART selects the best tree by testing its error
rate. In the holdout validation experiment, the CART 3.5. Comparisons of Alternative Methods
model achieves a top decile lift of 366 with the testing In the holdout validation experiment, BNs achieve the
data set. The resulting “optimal” tree with 379 nodes highest top decile lift (413), followed by latent class
(splits) is overwhelming, even for sophisticated users. regression (387), neural networks (376), and CART
In fact, it gives a set of weight scores to each of the (366). However, a single train-test validation is hardly
nodes of the tree, including all the recursive nodes. sufficient to assess the performance of a method. To
CART then ranks the variables in terms of their over- further compare the robustness of BNs against the
all explanatory power (weights) in classifying the other methods, we conduct a tenfold cross-validation
cases, with lifetime orders at the top (100), followed experiment (Kohavi 1995, Mitchell 1997). First, we use
by monetary value (78.3), recency (51.4), lifetime con- stratified random sampling to partition the data set
tacts (51.2), frequency (46.9), and order size (23.7). The into 10 disjoint subsets of equal size, that is, 10,629
other three variables make little difference. Despite cases with 5.4% of buyers. Following the standard
its apparent capability for classification, it is not easy practice of tenfold cross-validation, we then train and
even for the trained eye to understand how the results test all four methods 10 times using each of the
can help explain consumer responses. 10 subsets as the testing set (estimation sample) and
Latent class regression estimates a logit model all the remaining subsets combined as the training set
based on the assumption that the coefficients of the (validation sample).
predictors differ across unobserved latent segments, As is shown in Table 6, BNs provide the high-
and executes separate regressions for each of the est average lift in the top decile (408), followed by
latent classes. The procedure produces one set of latent class regression (401), ANNs (397), and CART
parameters for the predictor variables, including the (366). In the second decile, the ANNs have the high-
three residual variables that control for endogeneity, est cumulative lift (285), followed by BNs (283) and
and indicates their overall effects and their coeffi- CART (258), whereas latent class regression drops
cient estimates for each of the latent segments. In the to the last position (236). Overall, no method dom-
holdout validation experiment, the number of latent inates the others throughout the deciles. Although
classes that we tested ranged from two to five. It latent class regression provides the second-highest top
appears that a two-class model achieves the best fit, decile lift, its cumulative lift drops sharply in the sec-
based on the L-square statistic and the BIC value. This ond decile and starts to trail behind the other meth-
model achieves a top decile lift of 387 on the hold- ods, which indicates the instability of the method.
out sample. As with a typical regression, the overall In contrast, the results of BNs are rather stable,
coefficient estimates indicate the effects of the 12 pre- and their standard deviations of cumulative lifts are
dictor variables. In Table 5, Class 1 achieves a very mostly lower than those of the competing methods.
low R-square of 0.045 in comparison with Class 2 These results suggest that BNs not only predict con-
(R-square = 0712). In addition, Class 1 has different sumer response with a high level of accuracy, but also
Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
Management Science 52(4), pp. 597–612, © 2006 INFORMS 609

Table 6 Gains Table Based on Tenfold Cross-Validation methods achieve a reasonable level of classification
accuracy, their procedures are not as transparent to
Bayesian Latent class
networks ANNs CART regression the average users of marketing research, and the
Decile Cum. lift Cum. lift Cum. lift Cum. lift sources of improvement, if any, remain largely “hid-
den.” Their results are not easy to comprehend and
1 408.0 (17.0) 396.5 (24.3) 365.7 (18.4) 401.1 (26.3)
do not help the average user to gain insight into the
2 282.7 (8.2) 285.2 (12.8) 258.2 (15.2) 236.1 (9.3)
3 222.7 (4.8) 225.3 (6.7) 204.8 (8.9) 185.7 (8.2) problem. From the viewpoint of data mining, they are
4 186.8 (3.8) 190.6 (3.4) 175.5 (5.4) 156.5 (6.1) less intuitive in revealing how the predictor variables
5 162.2 (3.3) 166.0 (4.3) 149.0 (2.8) 135.8 (5.0) affect consumer response or what actions could be
6 144.5 (3.0) 147.5 (2.9) 132.5 (2.6) 120.6 (3.4) taken to improve the response rate.
7 129.8 (2.1) 131.7 (2.4) 120.1 (2.7) 107.0 (2.8)
In addition to its predictive accuracy, BNs exhibit a
8 118.0 (1.4) 119.7 (1.4) 110.9 (2.1) 95.5 (2.1)
9 107.4 (0.7) 108.9 (0.8) 104.2 (1.5) 93.4 (1.3)
high level of transparency in the optimization process.
10 100.0 (0.0) 100.0 (0.0) 100.0 (0.0) 100.0 (0.0) Furthermore, the results are easier to comprehend,
and provide more explanatory insight. BNs construct
Note. The reported figures are the means of the lifts of the 10 experiments,
and the standard deviations are given in brackets.
a graphic model to describe the relationships among
the directly observable variables, and then produce
the posterior and conditional probabilities among the
demonstrate a higher degree of robustness across dif- nodes to quantify their relationships. Whereas other
ferent data sets. methods such as latent class regression only describe
In Table 7, we compare these methods in terms of the conditional distribution of the dependent variable,
their transparency, interpretability, and insight. Over- BNs provide a joint probability distribution of all the
all, these methods are based on strong theoretical variables, including the conditional distribution of all
rationales and computational procedures, and they the exogenous variables. Based on the table of prob-
produce different results based on their model con- abilities, BNs can reveal the nonlinearities and inter-
figurations and the objective functions they emulate. actions among the variables, something that other
The ANN procedure produces the optimal structure methods cannot convey in a straightforward fashion.
by trial and error based on predictive accuracy, and Overall, the results of BNs are more interpretable for
its results specify the weights between the input vari- understanding the underlying relationships among
ables and the hidden nodes and between the hidden the variables, and provide more managerial insight
nodes and the output. The optimal number of latent into the problem and hence better decision support
classes in the latent class regression is determined by to managers in terms of customer selection for mar-
the L-square statistic and BIC value, which are rel- keting promotions. On a minor note, learning BNs
atively straightforward. Latent class regression pro- with EP is easy to execute and computationally effi-
duces parameter estimates for each of the predictors, cient, even with large databases. Whereas other meth-
which differ across the two latent clusters (classes). ods take hours to arrive at the optimal solution, BNs
In this sense, both methods rely on unobserved latent learned using EP take only a fraction of the time to
variables to improve their classification performance. converge.
The CART model specifies how the variables form a
splitting-tree model in a node-by-node fashion using 4. Conclusion
the recursive partitioning method (i.e., repeated use The results of this study show that BNs learned by
of the same variables in the tree), and then ranks the EP have performed satisfactorily against the objec-
variables in terms of their discriminating power (GINI tives of predictive modeling in direct marketing and
index). However, the tree structure is simply too com- data mining with large noisy databases, and that
plex for meaningful comprehension. Although these they have several advantages over the other methods.

Table 7 Comparison of the Alternative Methods

Method/criteria Bayesian networks ANNs CART Latent class regression

1. Model structure Graphic network model Network with input, output, hidden Recursive classification Logit model with latent classes
with nodes layer, and hidden nodes and regression tree
2. Results Posterior and conditional Weights of variables, nodes, and Weights of variables as Linear parameter estimates
probabilities biases splitting nodes
3. Accuracy∗ High High Medium High
4. Transparency High Low Low Medium
5. Interpretability/insight High Low Low Medium
∗
= based on the top decile lift.
 Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
610
First, BNs attain a high level of predictive accuracy,
and provide a good representation of the underlying
distribution of probabilities. Second, the optimization
algorithm of EP used in this study is straightforward.
The MDL metric as a fitness criterion balances model
accuracy and simplicity, and is effective in minimiz-
ing overfitting. Based on the MDL metric, the opti-
mization process by EP is completely tractable and
gives a clear indication of the relative performance of
the competing BN models identified in the evolution-
ary process. Together, they operate like a “white box”
with a great degree of transparency. Third, the ten-
fold cross-validation experiments indicate that BNs
deal with the bias-variance trade-off effectively, and
exhibit a high level of robustness.
More importantly, the BN topology is intuitively
appealing. A BN (a DAG) defines a qualitative model
for the joint probability distribution and describes
complex structures with efficiency and modularity.
The posterior probabilities help to determine the
effect of the predictors on consumer response and
reveal the nonlinear and interactive effects. Together
with the conditional probabilities, the resulting model
is able to elaborate how these variables together
explain consumer response and draw inferences
about consumer variables that have meaningful impli-
cations for managerial actions. These advantages pro-
vide a straightforward representation of the structure
of the problem, more interpretable results, and greater
explanatory insight that can provide decision support
(Chiogna 1997). It puts an intuitively appealing inter-
face on the machine learning process and demystifies
the methods of artificial intelligence.
The conventional approach to knowledge devel-
opment is largely theory driven. A researcher tests
hypotheses about the relationships among the vari-
ables of interest (Malhotra et al. 1999). Research with-
out a theoretical root is often considered to be lacking
in intellectual merit and analytical rigor, but the cur-
rent environment demands more problem-oriented
research and feasible methods to explore the vast
quantities of disaggregated data. Data mining aided
by sophisticated technology and versatile algorithms
can remove many of the restrictions associated with
traditional methods, and has become increasingly
important as a new way of discovering knowledge.
BNs together with EP can help researchers to gain
fresh insight into research problems by exploring rela-
tionships that were not anticipated, and provide a
viable alternative approach that can complement tra-
ditional methods. Given the increasing amount and
variety of data and the demand for reducing the
research cycle, these methods present efficient tools
for marketing managers to extract and update knowl-
edge in a timely fashion to assist decision making.
Management Science 52(4), pp. 597–612, © 2006 INFORMS
However, researchers need to be aware of the lim-
itations of these methods. BNs are discrete in nature.
Although discretization simplifies the learning pro-
cess and the resulting model, there may be a loss of
potentially useful information, and the model may
not fully capture all the details of the relationships.
EP can explore a wider search space to optimize BNs
by comparing many alternatives, but this process may
include invalid models and affect the efficiency of the
optimization process. Placing constraints on the learn-
ing algorithm based on existing domain knowledge
may help guide the search process to avoid invalid
models and to improve the overall efficiency and
accuracy. The integration of existing domain knowl-
edge using supervised learning is a fruitful avenue for
future research. Finally, the current BN model cannot
account for the element of firm behavior that exists in
this data set, and future studies should use a data set
that allows the explicit treatment of selection by the
management to correct for potential selection bias.
The convergence of the statistical approach and
machine learning methods represents one of the
most promising areas of research in the provision
of improved optimization solutions and better deci-
sion support for managers (Nakhaeizadeh and Taylor
1997). As a universal approximator, BNs can be
applied to a whole array of business problems.
BNs can handle multiple dependent variables and
deal with multiobjective optimization problems.
Given their efficiency in computing conditional and
marginal probabilities, BNs can also be used as a tool
for variable selection. BNs can be further enhanced
by tree-augmented networks (TAN) to perform deci-
sion tree analysis and to examine the effect of busi-
ness strategies. Latent class BNs can perform cluster
analysis for market segmentation and assist devising
differentiated promotion strategies. Obviously, how to
take advantage of these features of BNs requires more
applications and collaboration between management
researchers and machine learning specialists.
An online supplement to this paper is available on
the Management Science website (http://mansci.pubs.
informs.org/ecompanion.html).
Acknowledgments
The authors thank the associate editor and three anony-
mous reviewers for their insightful comments. They also
thank Dr. Guichang Zhang, Lin Li, Zhen Zhao, and
Yuanyuan Guo for their assistance in data processing and
conducting the experiments and Lingnan University for
funding this project.
Appendix. Conditional Probabilities of
Selected Variables
A complex DAG model like the one in Figure 3 has many
tables of conditional probabilities. We only show the follow-
ing ones as examples. As some tables have many entries,
only a few entries are shown here.
Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
Management Science 52(4), pp. 597–612, © 2006 INFORMS
A1: P (Purchase
Recency,
Frequency, Monetary value)
P (Purchase = x
Recency = 2,
Frequency = 4 Monetary value = 1
P (Purchase = x
Recency = 2,
Frequency = 4 Monetary value = 2
P (Purchase = x
Recency = 2,
Frequency = 4 Monetary value = 3
P (Purchase = x
Recency = 2,
Frequency = 4 Monetary value = 4
P (Purchase = x
Recency = 2,
Frequency = 4 Monetary value = 5
P (Purchase = x
Recency = 2,
Frequency = 4 Monetary value = 6
A2: P (Hcrd
Tele) x=0
P (Hcrd = x
Tele = 0 0.853
P (Hcrd = x
Tele = 1 0.365
A3: P (Frequency
Hcrd, Tele) x = 1 x = 2 x = 3 x = 4 x = 5
P (Frequency = x
Hcrd = 0, 0.211 0.510 0.159 0.105 0.015
Tele = 0
P (Frequency = x
Hcrd = 0, 0.026 0.162 0.238 0.447 0.128
Tele = 1
P (Frequency = x
Hcrd = 1, 0.217 0.351 0.200 0.194 0.038
Tele = 0
P (Frequency = x
Hcrd = 1, 0.065 0.298 0.255 0.304 0.078
Tele = 1
A4: P (Purchase
Frequency, Lifcont)
P (Purchase = x
Frequency = 2 Lifcont = 1
P (Purchase = x
Frequency = 2 Lifcont = 2
P (Purchase = x
Frequency = 2 Lifcont = 3
P (Purchase = x
Frequency = 2 Lifcont = 4
P (Purchase = x
Frequency = 2 Lifcont = 5

P (Purchase = x
Frequency = 5 Lifcont = 1
P (Purchase = x
Frequency = 5 Lifcont = 2
P (Purchase = x
Frequency = 5 Lifcont = 3
P (Purchase = x
Frequency = 5 Lifcont = 4
P (Purchase = x
Frequency = 5 Lifcont = 5
A5: P (Purchase
Frequency, Hcrd, Tele)
P (Purchase = x
Frequency = 1 Hcrd = 0 Tele = 0
P (Purchase = x
Frequency = 1 Hcrd = 0 Tele = 1
P (Purchase = x
Frequency = 1 Hcrd = 1 Tele = 0
P (Purchase = x
Frequency = 1 Hcrd = 1 Tele = 1

P (Purchase = x
Frequency = 5 Hcrd = 0 Tele = 0
P (Purchase = x
Frequency = 5 Hcrd = 0 Tele = 1
P (Purchase = x
Frequency = 5 Hcrd = 1 Tele = 0
P (Purchase = x
Frequency = 5 Hcrd = 1 Tele = 1
Note. Due to the symmetric nature of posterior probabilities, the
probability scores are omitted when purchase = 0.
611
References
Allenby, G. M., R. P. Leone, L. Jen. 1999. A dynamic model of
x=1
purchase timing with application to direct marketing. J. Amer.
Statist. Assoc. 94(446) 365–374.
0.001
Baesens, B., S. Viaene, D. van den Poel, J. Vanthienen, G. Dedene.
2002. Bayesian neural network learning for repeat purchase
0.194 modelling in direct marketing. Eur. J. Oper. Res. 138(1) 191–211.
Berger, P., T. Magliozzi. 1992. The effect of sample size and propor-
0.198 tion of buyers in the sample on the performance of list seg-
mentation equations generated by regression analysis. J. Direct
0.244 Marketing 6(1) 13–22.
Bhattacharyya, S. 1999. Direct marketing performance modeling
0.320 using genetic algorithms. INFORMS J. Comput. 11(3) 248–257.
Bitran, G., S. Mondschein. 1996. Mailing decisions in the catalog
0.421 sales industry. Management Sci. 42(9) 1362–1381.
Blodgett, J. G., R. D. Anderson. 2000. A Bayesian network model
of the consumer complaint process. J. Service Res. 2(4) 321–338.
Blundell, R. W., J. L. Powell. 2004. Endogeneity in semiparametric
x=1 binary response models. Rev. Econom. Stud. 71 655–679.
Chiogna, M. 1997. Probabilistic symbolic classifiers: An empirical
0.147 comparison from a statistical perspective. G. Nakhaeizadeh,
0.635 C. C. Taylor, eds. Machine Learning and Statistics. John Wiley &
Sons, New York.
Cooper, L. G. 2000. Strategic marketing planning for radically new
products. J. Marketing 64(1) 1–16.
Fogel, D. B. 1994. An introduction to simulated evolutionary opti-
mization. IEEE Trans. Neural Network 5 3–14.
Gönül, Füson F., Byung-Do Kim, Mengzhe Shi. 2000. Mailing
smarter to catalog customers. J. Interactive Marketing 14(2) 2–16.
Haddawy, P. 1999. An overview of some recent developments in
Bayesian problem-solving techniques. AI Magazine 20(2) 11–19.
Hansen, M. H., B. Yu. 2001. Model selection and the principle of
minimum description length. J. Amer. Statist. Assoc. 96(454)
746–773.
Haughton, D., S. Oulabi. 1997. Direct marketing modeling with
x=1 CART and CHAID. J. Direct Marketing 11(4) 42–52.
Heckerman, D. 1997. Bayesian networks for data mining. Data Min-
0.032 ing Knowledge Discovery 1 79–119.
0.050
Hu, M. Y., M. Shanker, M. S. Hung. 1999. Estimation of posterior
0.038 probabilities of consumer situational choices with neural net-
0.039 work classifiers. Internat. J. Res. Marketing 16(4) 307–317.
0.068 Jain, D., F. M. Bass, Y.-M. Chen. 1990. Estimation of latent class
models with heterogeneous choice. J. Marketing Res. 27(1)
94–101.
0.129
0.131 Jensen, F. V. 1996. An Introduction to Bayesian Networks. Springer-
Verlag, New York.
0.132
Kohavi, R. 1995. A study of cross-validation and bootstrap for accu-
0.135
racy estimation and model selection. Proc. 14th Internat. Joint
0.141 Conf. on Artificial Intelligence, Montreal, Canada.
Lam, W., F. Bacchus. 1994. Learning Bayesian belief networks—
x=1 An approach based on the MDL principle. Comput. Intelligence
10(3) 269–293.
0.018 Larrañaga, P., M. Poza, Y. Yurramendi, R. Murga, C. Kuijpers. 1996.
0.018 Structure learning of Bayesian network by genetic algorithms:
A performance analysis of control parameters. IEEE Trans. Pat-
0.018
tern Anal. Mach. Learning 18(9) 9.
0.018
Malhotra, N. K., M. Peterson, S. Bardi. 1999. Marketing research:
A state-of-the-art review and directions for the twenty-first
0.141 century. J. Acad. Marketing Sci. 27(2) 160–183.
0.141 Michie, D., D. J. Spiegelhalter, C. C. Taylor. 1994. Machine Learning,
0.130 Neural and Statistical Classification. Ellis Howard, New York.
0.130 Mitchell, T. 1997. Machine Learning. McGraw-Hill, New York.
Nakhaeizadeh, G., C. C. Taylor. 1997. Introduction. G. Nakhaei-
zadeh, C. C. Taylor, eds. Machine Learning and Statistics. John
Wiley & Sons, New York.
 Cui, Wong, and Lui: Machine Learning for Direct Marketing Response Models
612
Neal, R. M. 1996. Bayesian learning for neural networks. Lecture
Notes in Statistics, Springer, New York.
Pearl, J. 1988. Probabilistic Reasoning in Intelligent Systems: Networks
of Plausible Inference. Morgan Kaufmann, San Mateo, CA.
Pearl, J. 2000. Causality: Models, Reasoning, and Inference. Cambridge
University Press, Cambridge, UK.
Rao, V. R., J. H. Steckel. 1995. Selecting, evaluating, and updat-
ing prospects in direct mail marketing. J. Direct Marketing 9(2)
20–31.
Rissanen, J. 1978. Modeling by shortest data description. Automatica
14 465–471.
Smith, Richard J., Richard W. Blundell. 1986. An exogeneity test
for a simultaneous equation tobit model with an application to
labor supply. Econometrica 54(May) 679–685.
Management Science 52(4), pp. 597–612, © 2006 INFORMS
Venkatesan, R., V. Kumar. 2004. A customer lifetime value frame-
work for customer selection and resource allocation strategy.
J. Marketing 68(4) 106–125.
Warner, B. A. 1997. Bayesian learning for neural networks. J. Amer.
Statist. Assoc. 92 791–792.
West, P. M., P. L. Brockett, L. L. Golden. 1997. A comparative anal-
ysis of neural networks and statistical methods for predicting
consumer choice. Marketing Sci. 16(4) 370–391.
Wong, M. L., W. Lam, K. S. Leung. 1999. Using evolutionary com-
putation and minimum description length principle for data
mining of Bayesian networks. IEEE Transactions Pattern Anal.
Mach. Intelligence 21(2) 174–178.
Zahavi, J., N. Levin. 1997. Applying neural computing to target
marketing. J. Direct Marketing 11(4) 76–93.