Abstract
Editorial Measuring the impact of personalization and recommendation on user behaviour In recent years, several methods for generating and presenting personalized information have been developed. They are aimed at taming the information overload caused by the vast range of continuously growing content available on the World Wide Web. Examples include adaptive hypertext, which adapts user interfaces to its users' interests and needs (e.g. in mobile portals), and recommender systems that try to predict, for each user, which items, e.g. news, movies or travels are relevant. Recommender systems use various artificial intelligence (AI) and information retrieval (IR) techniques to build their predictions, and they typically interact with the user for presenting and revising their recommendations (Ricci et al., 2010). Many of these techniques and systems have made their way from research into commercial applications and are now widely used in major eCommerce portals. At the same time, new techniques are being proposed, improving recommendation effectiveness or offering new ways for users to participate, such as social networks on Web 2.0 platforms. The assessment of the benefits, for the end users and the service providers, brought by the scientific and technological development of the personalization techniques is obviously an important issue. In fact, choosing the right evaluation method for personalized web applications, identifying the influential success factors behind different techniques or interpreting results coming from online experiments remain largely open research issues. Off-line experiments, which estimate the recommendation prediction error using existing recommendation datasets (as described in the seminal article of Herlocker et al., 2004), is the traditional state-of-practice for evaluating recommender systems. A small survey of the evaluation designs adopted in the research articles dealing with the topic of recommender systems, which appeared in the ACM Transactions on Information Systems over a 5 year period (2004–2008), showed that three quarters of these studies instrumented off-line experiments and half of them chose movie recommendations as their application domain (Jannach et al., 2010). In fact, many researchers have stressed the limitations of off-line evaluations, to the point that some have argued that the true accuracy of a recommender system can never be directly measured using these approaches. However, the widespread use of recommender systems makes it crucial to develop evaluation approaches that can realistically and accurately assess the true effectiveness of the recommendations and their effects on the users. For these reasons, this special issue contains five selected articles that focus on the understanding of how persona-lization and recommendation impact on the user expectations , beliefs and behaviour during and after interaction. The articles not only consider traditional websites, but also reach out into mobile and ubiquitous communication and interaction scenarios. Furthermore, the selected articles focus on the impact of personalization and recommendation in diverse application domains, ranging from traditional e-commerce and news content personalization to mobile, in-vehicle and interactive TV interfaces. Overview of the papers. Selecting the recommendation approach that is best suited to a specific application or product domain is an open research issue, and only a few works have discussed the topic of matching recommendation technologies with application domains (Burke and Ramezani, 2010). In fact, few empirical studies have compared different recommendation strategies in different application settings to gain comparative evidence for deciding between different system designs (Xiao and Benbasat, 2007). The article by Ochi et al. (2010) therefore constitutes one piece of additional works for building a general theory on the design parameters of recommender systems in specific application domains. They conducted an empirical study under lab conditions where a 2 (content-based vs. collaborative recommendation approaches) Â 2 (few vs. many feature dimensions used to generate recommendations) Â 2 (experience vs. search product type) mixed-design experiment was conducted. Users were initially asked to specify their profile by filling out a questionnaire with either social or content-related questions and were then presented with a recommendation list. In order to control the effects of the recommendation list itself, all participants were presented with exactly the same recom-ARTICLE IN PRESS www.elsevier.com/locate/ijhcs