214 J. Buder, C. Schwind/Computers in Human Behavior 28(2012)207-216 Table 2 Contrast bet mmerce requirements and educational requirements with regard to recommender system design, and resulting design strategies for educational contexts. Issue Requirement for e- Educational requirement Design strategy for educational recommender systems commerce Recipient role/system- Personalization with Context-aware personalization with regard Hybrid systems regard to taste to knowl ledge/ac tivities Learner feedback Recipient role/social P preference-consistency th Persuasiveness; challenges and critical strong arguments: expertise/authority cues: preferen adaptation Producer role/system-Low burden: implicit Participation, metacognitive stimulation Explicit ratings centered adaptation elicitation Producer role/social High number of ratings High number of ratings Providing utility information adaptation Dakof, 1990). Moreover, providing utility information is more importantly, user activity should be minimized, e.g. by implicit likely to appeal to the collaborative spirit that fuels recommender g hybrid systems with ontol systems. This strategy of minimal s, though it can be argued th ated in an age 4. Conclusions This paper explored the potentials of personalized jected, or ns nder systemsDakof, 1990). Moreover, providing utility information is more likely to appeal to the collaborative spirit that fuels recommender systems. 4. Conclusions This paper explored the potentials of personalized recom￾mender systems in educational settings. It is argued that recom￾mender systems fit nicely to important principles in the learning sciences: (1) Recommender systems are peer technologies that shift responsibility away from dedicated experts. (2) Recom￾mender systems are technologies where the quality of content is not traceable to any individual output, but rather to the collective behavior of a community. (3) Recommender systems provide user control, thereby facilitating self-regulated, exploratory, and auton￾omous learning. (4) Recommender systems provide guidance to learning activities. (5) Recommender systems are adaptively tai￾lored to the needs and requirements of learners. However, it would clearly be a mistake to apply standard rec￾ommender systems to learning scenarios without adapting them to educational needs. Section 3 of this paper structured issues sur￾rounding educational applications of recommender systems with regard to two roles that learners exhibit, viz. as recipients of infor￾mation and as producers of data. For each of these two roles, two issues were discussed: One with regard to system-oriented adapta￾tions that enable proper functioning of educational recommender systems, the other with regard to social adaptations that exert an influence on how learners react to and act upon recommendations. On the basis of theoretical and empirical findings from various re￾search fields design-related questions were posed and answered. A summary of that discussion can be found in Table 2. The leftmost column of this Table represents the four issues that were raised in the discussion. The second and third columns contrast the requirements for classical recommender systems in e-commerce vs. educational recommender systems. And the rightmost column proposes design strategies for educational recommender systems. Rather than repeating the issues discussed in preceding sec￾tions, we’d like to point out some recurring thoughts, particularly about the differences between e-commerce recommender systems and educational recommender systems. In classical e-commerce scenarios, the main goal of designers is to increase cross-sell of items. The corresponding recommender systems can be character￾ized by a number of typical features. These systems are adapted to user taste, and they try to keep the potential burden of system usage to a minimum. As a consequence, the solutions rely heavily on constraining system output, preferably through machine intel￾ligence. Although there is some awareness among designers that recommender systems should support serendipity and diversity of result lists (Ziegler, McNee, Konstan, & Lausen, 2005), the gen￾eral consensus seems to be that a recommender system should not come up with anything that is unexpected by a user. Most importantly, user activity should be minimized, e.g. by implicit preference elicitation, or by employing hybrid systems with ontol￾ogies and user modeling techniques. This strategy of minimal interference might be useful in commercial contexts, though it can be argued that it appears as somewhat outdated in an age where user-generated content has become so pervasive. However, it is quite evident that such a strategy should not be adapted for learning contexts. Educational recommender systems are not geared at selling items, but at facilitating learning. Learning is an active and constructive process (Vygotsky, 1978), therefore it seems only natural to involve and engage learners in the very pro￾cess they undergo. This is reflected in the suggested design strate￾gies of Table 2: It can be helpful both for system performance and for learner satisfaction to provide customization options, and to give opportunities to leave feedback on system accuracy. Rating of items should not be regarded as a burden to learning, but rather as an opportunity for learning. Making the community visible by emphasizing the utility that rating has for others fuels the collab￾orative spirit that is needed for effective recommender systems. And finally, educational recommender systems should involve learners by challenging them. Rather than providing learners with a strongly constrained environment, they should leave ample room for exploration and confront learners with unexpected content, thus allowing for serendipity and learning through discovery. All these differences between e-commerce and educational contexts call for an adaptation of personalized recommender systems so that they can become powerful tools for learning. Of course, this overview of the potentials of recommender sys￾tems in educational contexts is not exhaustive. For instance, we did not cover interface issues, basically because we think that they do not require specific adaptations for educational scenarios. Readers who are interested in these aspects might refer to usability studies (Herlocker, Konstan, & Riedl, 2000), studies showing serial position effects for recommended items (Felfernig et al., 2007), or work on the display of ratings (Cosley, Lam, Albert, Konstan, & Riedl, 2003). Our overview was also restricted to the impact of recommenda￾tions during learning, and we did not address general aspects of recommender use, most of which is covered by literature on trust (Swearingen & Sinha, 2002). Finally, we focused on the most com￾mon types of recommender systems, thereby excluding variations like recommender systems for groups (Buder & Schwind, 2011), or people recommender (Cai et al., 2011). Many of the design considerations that were suggested in this paper rest on speculation. This is due to the scarcity of research in a relatively new field. We hold that research on personalized recommender systems should not only focus on the development of better algorithms and implementations, but should be comple￾mented by sound, empirical work on how learners react to and act upon recommender systems. In order to see whether the assumptions of this paper can be confirmed, must be rejected, or require refinement, three types of research are needed. First, we need more practical implementations of recommender systems Table 2 Contrast between e-commerce requirements and educational requirements with regard to recommender system design, and resulting design strategies for educational contexts. Issue Requirement for e￾commerce Educational requirement Design strategy for educational recommender systems Recipient role/system￾centered adaptation Personalization with regard to taste Context-aware personalization with regard to knowledge/activities Hybrid systems Learner feedback Customization Recipient role/social adaptation Persuasiveness; preference-consistency Persuasiveness; challenges and critical thinking Strong arguments; expertise/authority cues; preference￾inconsistent recommendations Producer role/system￾centered adaptation Low burden; implicit elicitation Participation, metacognitive stimulation Explicit ratings Producer role/social adaptation High number of ratings High number of ratings Providing utility information 214 J. Buder, C. Schwind / Computers in Human Behavior 28 (2012) 207–216