No abstract

The author is grateful to Scott Neslin and the two anonymous JMR reviewers for their careful reading, detailed comments, and insightful suggestions on multiple versions of this article. They made this article vastly better than it otherwise would have been. ANAND V. BODAPATI* Many firms use decision tools called "automatic recommendation systems" that attempt to analyze a customer's purchase history and identify products the customer may buy if the firm were to bring these products to the customer's attention. Much of the research in the literature today attempts to recommend products that have a high probability of purchase (conditional on the customer's history). However, the author posits that the recommendation decision should be based not on purchase probabilities but rather on the sensitivity of purchase probabilities to the recommendation action. This article attempts to model carefully the role of firms' recommendation actions in modifying customers' buying behaviors relative to what the customers would do without such a recommendation intervention. The author proposes a simple consumer behavior model that accommodates a transparent role for a firm's recommendation actions. The model is expressed in econometric terms so that it can be estimated with available data. The author studies these ideas using purchase data from a real e-commerce firm and compares the performance of the proposed main model with the performance of benchmark models. The author shows that the main model is better than benchmark models on key measures.

The success of Internet social networking sites depends on the number and activity levels of their user members. Although users typically have numerous connections to other site members (i.e., “friends”), only a fraction of those so-called friends may actually influence a member's site usage. Because the influence of potentially hundreds of friends needs to be evaluated for each user, inferring precisely who is influential—and, therefore, of managerial interest for advertising targeting and retention efforts—is difficult. The authors develop an approach to determine which users have significant effects on the activities of others using the longitudinal records of members' log-in activity. They propose a nonstandard form of Bayesian shrinkage implemented in a Poisson regression. Instead of shrinking across panelists, strength is pooled across variables within the model for each user. The approach identifies the specific users who most influence others' activity and does so considerably better than simpler alternatives. For the social networking site data, the authors find that, on average, approximately one-fifth of a user's friends actually influence his or her activity level on the site.