16 211 ion)(Ent-2003). Moreover, involving learners into the very process of recom- mendations caters well to the spirit of learning as a constructive collaborative activity. This leads to our first conjecture: for re are in orde g activities. rT itiit h -osdo arners in of these two strategies ld be that learner nd affective reactions e design of lars hold nity to fee mendatior strategy is le n edu- hybrid rec strong ner involver ner control higher satisfac retzel,2011).Fororientation, achieving orientation, and meaning orientation) (Ent￾wistle, 1988) might require different recommendations; recom￾mendations might ideally take metacognitive skills and strategies (Weinstein & Mayer, 1986) into account; and they should be adapted to the goals of a learner (Boekaerts, 1998) – for instance, a learner who wants to find an explanation on a specific algebraic problem needs different recommendations than a learner who seeks for general resources on algebra. As Drachsler, Hummel, & Koper (2009) maintained, educational recommender systems would ideally be able to situationally identify those items that cor￾respond to a learner’s zone of proximal development (Vygotsky, 1978), the level of ability that the learner is able to master through scaffolding. In other words, in order to adapt to learner knowledge and learning activities, recommender systems must be context￾aware (Adomavicius & Tuzhilin, 2011). The classical approach of collaborative filtering through the analysis of simple ratings might not be very helpful, as a high rating could mean that a learner found an item easy, or challenging, or fun. However, there are dif￾ferent ways to achieve context-awareness. A first strategy to make recommender systems context-aware is to use machine intelli￾gence, e.g. through so-called hybrid recommender systems (Burke, 2002). These combine recommender algorithms like collaborative filtering with content-based filters and/or learning modeling tech￾niques (Brusilovsky, 2001). By way of our digital library example from Section 2, a publication recommender system could use a hy￾brid model that combines rating information with metadata. For instance, if our student gave a high rating for an article, the system could automatically increase the probability that other publica￾tions from the same author are recommended. Ontologies, tags and metadata can be used to describe learning items in more de￾tail, and modeling techniques can be used to describe learners in more detail. Having ontologies can also help to address the prob￾lem of sequential dependencies among learning items, and they might pave the way for systems that do not recommend isolated items, but actual learning paths (Drachsler, Hummel, & Koper, 2009). As to date, there are few examples of hybrid educational recommender systems that go beyond a prototype development, let alone a full system evaluation. However, in a detailed computer simulation study, Nadolski et al. (2009) found that different types of recommender systems yielded much better results (graduation percentages, user satisfaction, graduation times) than no recom￾mendations. Further, the authors found that hybrid recommender systems outperformed purely rating-based and purely ontology￾based recommender systems, although not by a significant margin. A real-world investigation of hybrid educational recommender systems compared a group using a hybrid personalized recom￾mender system for learning activities with a no-recommendation control group (Drachsler, Hummel, van den Berg, et al, 2009). In a usage study covering 4 months, they found that groups using the recommender system did not complete more activities, but completed them faster, exhibited a greater variety of learning paths, and expressed higher satisfaction. These examples show that hybrid educational recommender systems are likely to have measurable effects on learning-related variables. A second strategy to increase context-awareness does not rely on machine intelli￾gence, but on involving learners into the recommendation process. For instance, learners could choose among different learning paths depending on their learning styles or concrete learning goals. Moreover, dialogs could be provided that give learners an opportu￾nity to feed back on the situational adequacy of received recom￾mendations. While it appears that the learner involvement strategy is less popular among system designers than the use of hybrid recommender systems, it should be noted that active lear￾ner involvement might have additional benefits. For instance, lear￾ner control and customizability of system output are related to higher satisfaction and trust (McNee, Lam, Konstan, & Riedl, 2003). Moreover, involving learners into the very process of recom￾mendations caters well to the spirit of learning as a constructive and collaborative activity. This leads to our first conjecture: 3.1.1.1. How to achieve system-centered adaptation for recipi￾ents? Recommender systems must be context-aware in order to correctly diagnose learner knowledge and learning activities. This can be accomplished either through machine intelligence (hybrid recommender systems) or through involvement of learners into the recommendation process itself (customization, feedback loops). It is an empirical question which of these two strategies is superior, but a tentative conclusion could be that learner involvement has additional educational benefits. 3.1.2. Social adaptation Information processing has a social dimension: It is colored by attitudes, judgments, stereotypes, and affective reactions (Bandura, 1986). This lends a social dimension to the design of educational recommender systems as well. Some scholars hold that computers are social actors (Nass, Moon, Morkes, Kim, & Fogg, 1997), or that they are persuasive technologies that can exert social influence (Fogg, 2003). More specifically, the idea that recommender systems are perceived as social actors is sup￾ported by the observation that users ascribe a personality to them (McNee, Riedl, & Konstan, 2006). Personalized recom￾mender systems mimic a knowledgeable person, a person that does not only have information about a huge number of items, but also about the tastes and preferences of a user. However, we do not always follow a recommendation by a human being, and of course the same might apply to recommendations from a recommender system. This raises questions about the condi￾tions under which the selection of recommended items can be influenced. In order to answer these questions, it is helpful to re- flect on biases in human information processing. Some of these biases are conducive to learning and can be put to good use by making recommendations more appealing. Other biases in information processing are detrimental to particular types of learning, so recommender systems should be designed to over￾come these biases. We now turn to conducive biases in the con￾text of literature on persuasion, followed by detrimental biases in the context of selective exposure literature. Dual-process models of persuasion have outlined the boundary conditions that determine whether people are more or less in￾clined to follow a persuasive message such as a recommendation. According to the elaboration likelihood model (Petty & Cacioppo, 1986), the degree to which a persuasive message is elaborated de￾pends on a recipient’s motivation and ability to process the mes￾sage. Low personal relevance of the message topic undermines motivation, whereas distraction during processing impedes ability. If motivation and ability are high, messages are carefully scruti￾nized, and persuasion mainly depends on so-called message char￾acteristics; in contrast, if motivation and/or ability are low, persuasion mainly depends on so-called source characteristics (McGuire, 1969). As motivation and ability are not directly control￾lable, design of educational recommender systems should try to unfold persuasive power through message characteristics and source characteristics. As to message characteristics, the variable that is most often associated with them is argument strength. For instance, the elaboration likelihood model predicts that under conditions of high elaboration (high motivation and ability), a strong argument becomes persuasive, whereas a weak argument is likely to be rejected. As a consequence, the item pool of an edu￾cational recommender system should contain as many strong arguments as possible. A second way to influence the persuasive￾ness of a recommender system is through source characteristics, i.e. perceived attributes of a sender (Yoo & Gretzel, 2011). For J. Buder, C. Schwind / Computers in Human Behavior 28 (2012) 207–216 211