ACM SIGKDD 2006 Tutorial
Data Analytics for Marketing Decision Support

August 20th, 2006, Philadelphia, PA

Tutorial Co-Organizers: Saharon Rosset and Naoki Abe

Introduction Today, with increasing amounts of business data accumulated within enterprises and wider availability of data management infrastructure, the use of analytical tools to support marketing decisions is becoming commonly acknowledged as key to success in a wide range of business areas. One particularly important challenge one faces when delivering marketing decision support systems is that of understanding the true value of customers, beyond the currently observed value. Some key issues that arise when attempting to meet this challenge are:

Customer targeting for the so-called whitespace, i.e., those entities that are currently non-customers of the enterprise in question.
Customer lifetime value estimation, i.e., estimating the expected value of a customer over the course of its long term interaction with the enterprise.
Customer wallet estimation, i.e., estimating the potential needs, and hence value, of a customer, inclusive of the portion currently spent on purchases from the competitors.

As the expectation and dependency on data mining tools grows in marketing support practices, so do the challenges and issues, and the gap widens between what is solvable by straightforward applications of analytical methods and what really needs to be solved for the business problem at hand. This tutorial aims at clarifying the nature of this gap, and describing what solutions there are to bridge it. The tutorial will consist of three parts. First, we will give an overview of the type of challenges that are often faced by data mining practitioners when applying analytic tools to marketing decision support problems. We will then review a number of general methodologies that can be used to address and partly solve these problems, either by solving them under restrictive assumptions, or by requiring further input for quantities that cannot be estimated from the data. Finally, we will describe in detail a number of case studies, in order to illustrate how one goes about designing a practically useful solution to a real world marketing decision support problem.

The Challenges In marketing decision support problems, one wishes to use data analytics to help obtain better understanding of the effects of marketing actions, and ultimately, make better marketing decisions. The data on which analytics are applied, however, may not contain all necessary information. In particular, the effect of marketing actions on future, potentially long-term, behavior of a customer is not directly observed, and the behavior of the customer beyond its interaction with the enterprise collecting the data is typically unobserved. (Some exceptions exist to this rule, as is the case, for example, with credit rating firms). This mismatch between the goal of analytics and the data is the source of much of the challenge that we wish to address in this tutorial. Specifically:

The data may not contain information on the decisions/actions that one is trying to optimize, thereby making it difficult to estimate the effect of decisions/actions from the data.
The data may not contain direct information on the objective of marketing decision making, such as customer lifetime value and customer wallet.
The data may have been collected in a different context and thus straightforward application of estimation methods would result in biased estimation.
Correlations observed in the data may not necessarily correspond to causality, and hence the results of analysis do not give rise to actionable rules for marketing optimization.

There are many other issues in applying data mining to business analytics in general, such as the issues of data quality and cleansing and that of stationarity of distributions. In this tutorial, we focus on the challenges listed above, which are specific to the goal of supporting optimal marketing decision making.

Methodologies These problems have been encountered by many data mining experts, researchers and practitioners, and solutions exist with varying degrees of sophistication and effectiveness. Some methodologies that are of direct relevance include:

Markov Decision Processes (MDP) and reinforcement learning methods, which resolve credit assignment of marketing actions to their delayed effects, and hence allow modeling of long-term effects of those actions.
Bayesian Networks, which formalize constraints on possible causal effects by setting a network structure. This approach allows us to make optimal use of available pieces of information to make inference on unobserved quantities, such as the customer wallet.
Integration of expert knowledge and data analytics, allowing human experts to specify parameters that cannot be learned purely from available data. Bayesian methods can create a smooth combination, where priors are specified that can be ``over-ridden'' by evidence from data, if available.

We plan to review in some detail these approaches and their effectiveness in the present context of marketing decision support problems. We will also touch upon some of the other methodologies that are applicable to this domain, when we widen the scope of issues and problems: %albeit to lesser extent to the specific problems illustrated above:

Controlled experiments, where the possible business actions are played out on appropriately chosen sub-populations, to give ``unbiased'' evaluation of the success of these actions. In particular, active learning methods allow the dynamic design of controlled experiments to more efficiently collect data for analysis.
Bias correction methods, which allow for correction of biases introduced by data sampled in a different context.
Cost sensitive learning methods, which allow the cost aspect of the marketing problem to be considered, and better guide the estimation process and outcomes.

Case Studies As noted in Introduction, correct application of data mining technology is critical to the success of marketing support systems and practices. In this tutorial, we will select at least three case studies based on our experience in this area, to illustrate this point. We will describe in detail the challenges faced and methods applied in the following real world applications (we will probably choose three of these, as time dictates):

Application of the Markov Decision Processes (MDP) approach to address customer life time value modeling and marketing optimization in the retails industry.
Modeling of unobserved customer wallets through a couple of different approaches: quantile regression and integration of domain knowledge via efficient Bayesian network inference.
Application of data-enhanced modeling techniques to the problem of optimizing sales targeting policies for acquiring ``white space'' customers for software products.
Integration of expert knowledge and data mining to model the effect of marketing actions on customer life time value in a telecom environment.