Anand V. Bodapati
To implement the common customer relationship management notion of add-on selling, many firms use decision tools called “automatic recommendation systems” that attempt to analyze a customer’s purchase history and to identify products the customer might buy if the firm were to bring these products to the customer's attention. This article addresses the question of how customer histories should be analyzed to identify products for the firm to recommend to a specific customer. This article differs from others in the literature in three important ways. First, articles in the literature today attempt to recommend products that have a high probability of purchase (conditional on the customer’s history). However, the recommendation decision should be based not on purchase probabilities but rather on the sensitivity of purchase probabilities to the recommendation action. This is a fundamentally important framing that has been absent in the literature. This article attempts to model the role of firms’ recommendation actions in modifying the customers’ buying behaviors relative to what they would do without such recommendation intervention. It is not possible to make optimal recommendations to the customer without an explicit understanding of the effect of those recommendations on purchase behavior. Thus, the author proposes a simple consumer behavior model that accommodates a transparent role for the firm’s recommendation actions. The model is expressed in econometric terms so that it can be estimated with available data.
Second, this article assumes that the firm has access only to purchase behavior data. In contrast, the overwhelming majority of existing articles assume that firms have access to ratings data from customers (e.g., each customer rates each item on a five-point scale to indicate the extent to which he or she likes or dislikes an item).
The third aspect of differentiation in this article is the careful attention given to scalability and efficiency of the estimation algorithms in settings in which the firm may have millions of customers and products. The author shows that the key likelihood function optimization is equivalent to a sequence of operations minimizing the weighted Frobenius norm with respect to a certain target matrix, which in turn can be expressed as a sequence of eigenvalue–eigenvector calculations. The consequence of this is that maximum likelihood estimates can be obtained with relative ease and high speed. The author studies these ideas using purchase data from a real e-commerce firm. The author compares the performance of the main model proposed in the article with the performance of benchmark models. He shows that the main model is better than the benchmark models on key measures.
Anand V. Bodapati is Associate Professor at the Anderson School of Management, University of California, Los Angeles. His research interests are in the areas of customer relationship management, direct marketing, and the translation of consumer psychology findings to econometric models and in methodological issues related to assessing customer responsiveness to marketing. He received his doctorate degree from Stanford University.
Journal of Marketing Research, Vol. XLV, No. 1, February 2008
View Table of Contents.