Dissertation/Thesis Abstract

Two essays on “mining market basket data: Models and applications in marketing”
by Li, Xiaojun, Ph.D., The George Washington University, 2008, 80; 3315049
Abstract (Summary)

A retailer typically makes decisions on what products to promote, how, and when. These decisions, often referred to as marketing mix activities, are especially important given that pricing and promotion decisions in one category can not only influence sales in the promoted category, but also in other categories. For example, a price reduction in cake mix may not only boost its own sale but also sales of cake frosting. It would thus be in the retailer's interest to coordinate the marketing activities across products in multiple categories so as to maximize profit. One such model for maximizing store profit could be stated as follows. Suppose there are n products. Let the decision variables that relate to the pricing and promotion decisions for each product in week t be represented by pt = {p1 t, ...pnt} and Promo t = {Promo1t, ..., Promont}, respectively. Clearly, given the presence of cross-category effects, the sales of product i during week t could be expected to be a function of Promot and p t, resulting in the following expression for the revenue Rt during week t, [special characters omitted]

where sit is the sales of product i at week t and pit is the price of product i at week t. The decision problem then is to determine values for the Promo t and pt variables such that the revenue in equation 1 is maximized.

Before one can address such a goal, two pieces of critical information are needed. First, given the large number of products carried in a store, one needs to know what products are associated. Second, one needs to understand how promotions work across categories. Since the 1990s, advances in information technology have made the collection and storage of consumers' purchase history and shopping basket content technically and economically feasible. Such market basket data contain valuable information about product association and promotion effects, and make possible the analysis of coordinated marketing mix activities. This dissertation consists of two essays in data mining models of market basket data.

Essay 1 is titled "Market basket analysis using Bayesian Networks". This essay addresses the question of how promotions work across categories. Promotions in one product category can affect sales of products in another category either directly or indirectly. Given a set of product categories and market basket data, we analyze the presence of cross category impacts using Bayesian Networks. We model the occurrence of a product category, and not the number of units (of a product category) in a basket. The data set we employ is an IRI market basket data set that contains transactions including 22 categories over 2 years for 500 panelists. Bayesian Networks are learned from this data and are used to identify the underlying dependencies across product categories. Specifically, we study how the associations across categories vary based on marketing mix activities, and also based on demographics. The results from such an analysis can help in (1) identifying clusters of categories wherein associations exist primarily between categories within a cluster and not across clusters, and (2) in making predictions on basket choices given a set of specific marketing mix activities. The ability of Bayesian networks to learn based on new evidence also makes such an approach possible in an online context when customers' choices can be observed, and marketing activities can be dynamically customized.

Essay 2 is titled "Localized rule discovery in market basket data" and builds on traditional association rule mining algorithms to identify pairs of products that are associated. Due to consumer heterogeneity, the association between products may vary across consumer segments. Two products can be globally associated or locally associated. For the latter, associated products are matched with consumer segments within which the localized associations are strong. We illustrate the algorithm on the same shopping market basket data set used in Essay 1. The results from this analysis should help the retailer in identifying customer segments in which specific pairs of products are strongly associated, and also help determine the marketing mix effects on cross product associations.

Indexing (document details)
Advisor: Prasad, Srinivas, Rau, Pradeep
Commitee: Soyer, Refik
School: The George Washington University
Department: Business Administration
School Location: United States -- District of Columbia
Source: DAI-A 69/06, Dissertation Abstracts International
Subjects: Marketing
Keywords: Bayesian, Data mining, Market basket, Product association, Promotions
Publication Number: 3315049
ISBN: 9780549637288