Optimizing the connectedness of recommendation networks for retrieval accuracy and visiting diversity of random walks

Recommendation networks have been widely implemented on online systems, where each object connects to several similar others with hyperlinks. A typical example is Amazon’s “customers who bought this item also bought” list. Though the recommendation list length is potentially pivotal in determining the connectedness of recommendation networks, how such list length influences the systems’ navigation efficiency has not been well explored. The present paper measures the accuracy of users’ short-term surfing on recommendation networks in retrieving historical interests, and the diversity of visited objects during such surfing. Analytical results based on three empirical datasets and three similarity algorithms show that, providing more recommendations in the list can promote the diversity of objects that the users could potentially visit. However, these recommendations may also divert users’ attention from the most relevant ones, leading to decreased retrieval accuracy. To achieve the best accuracy, recommendation lists need to be relatively short, depending on the expected surfing duration and the applied similarity algorithms. Consequently, this study uncovers the profound impact of recommendation network connectedness on the short-term surfing accuracy and diversity, and thereby highlights the necessity of tailoring the often-overlooked recommendation list length for enhanced navigation efficiency.