Distributed Recommender Systems for Internet Commerce

Abstract

Recommender systems (RSs) present an alternative information-evaluation approach based on the judgements of human beings (Resnick & Varian, 1997). It attempts to automate the word-of-mouth recommendations that we regularly receive from family, friends, and colleagues. In essence, it allows everyone to serve as a critic. This inclusiveness circumvents the scalability problems of individual critics—with millions of readers it becomes possible to review millions of books. At the same time it raises the question of how to reconcile the many and varied opinions of a large community of ordinary people. Recommender systems address this question through the use of different algorithms: nearest-neighbor algorithms (Resnick, Iacovou, Suchak, Bergstrom, & Riedl, 1994; Shardanand et al., 1994), item-based algorithms (Sarwar, Karypis, Konstan, & Riedl, 2001), clustering algorithms (Ungar & Foster, 1998), and probabilistic and rule-based learning algorithms (Breese, Heckerman, & Kadie, 1998), to name but a few. The nearest-neighbor-algorithm-based recommender systems, which are often referred to as collaborative filtering (CF) systems in research literature (Maltz & Ehrlich, 1995), are the most widely used recommender systems in practice. A typical CF-based recommender system maintains a database containing the ratings that each customer has given to each product that customer has evaluated. For each customer in the system, the recommendation engine computes a neighborhood of other customers with similar opinions. To evaluate other products for this customer, the system forms a normalized and weighted average of the opinions of the customer’s neighbors.