Comparing and modeling the use of online recommender systems

Abstract

This study explores a new way to model the adoption of AI, specifically online recommender systems. It aims to find factors that can explain the variation in usage in terms of differences between individuals and differences over technologies. We analyzed survey data from users of online platforms in the U.S. using a two-level structural equation model (SEM) (N = 1007). In this model, the dependent variable was the usage rate, which was defined as the share of time a person used a particular recommender system (e.g., “People You May Know”) when they use the platform (e.g., Facebook). The individual responses (within-systems level) were clustered in the 26 recommender systems (between-systems level). We hypothesized that three technology-specific factors, adapted from the Diffusion of Innovations (DOI) theory and the Unified Theory of Acceptance and Use of Technology 2 (UTAUT2), could explain the variations in usage at both levels: perceived performance expectancy (PE), perceived effort expectancy (EE), and perceived hedonic motivation (HM). Our estimated model showed that usage was associated with PE and HM at the within-system level and only with PE at the between-system level. A considerable part of the variation in usage across the 26 systems could be explained by PE only (R2 = 0.30). The most important contribution to practitioners is that this study provides evidence for the idea that there are inherent, measurable differences across recommender technologies that affect their usage rates, and specifically it finds usefulness to be a key factor. This is potentially valuable for app developers and marketeers who look to promote the adoption of novel recommender systems. The main contribution to the literature is that it presents a proof-of-concept of a two-level model for AI adoption, conceptualizing it as an effect of both variations over users and variations over applications. This finding is potentially valuable for policymakers, as better predictive models might enable improved assessments of AI's social implications. In future studies, the two-level approach presented here could be applied to other forms of AI, such as voice assistants, chatbots, or Internet of Things (IoT).