213 comparisons(Festinger, 1954). This was investi- 007). They gave bogus sar Bre through ra tional rec 3.2.1.1.Hov ers? Recon tation metl between par itation m Through rat merits of able metac investigate gh so- mative parisons or th thers. identity(Ta of rat- and this adher s both ymous(Reich assoch about van Yperen,methods is that users might be better able than computers to make judgments on subjective and fuzzy rating categories (Norman, 1993). For instance, in the digital library example it would be dif- ficult for implicit elicitation methods to differentiate those publica￾tions that our student selected and subsequently regarded as inappropriate from those publications that she found useful after reading the abstract. This problem would not have occurred if both the inappropriate and the useful item had been explicitly rated. Recommender systems in e-commerce have the primary goal to make users aware of items that might be interesting to them, thereby increasing cross-sell of items. In this light, it is quite rea￾sonable not to burden users with additional rating activities and opt for implicit preference elicitation. In contrast, for educational recommender systems the use of ratings might have additional benefits, as explicit elicitation can be regarded as a form of partic￾ipation (Ling et al., 2005). Participation is a highly pervasive notion in the learning sciences. For instance, the amount of contributions that a learner produces during interaction is regarded as an impor￾tant antecedent of learning outcomes (Cohen, 1994). Participation requires learners to reflect on an issue, thereby leading to deeper elaboration (Kollar et al., 2006), and in this regard explicit rating instructions can serve as a valuable metacognitive prompt (Palinc￾sar & Brown, 1984). For this reason, explicit elicitation of user data through ratings appears to be a promising approach for educa￾tional recommender systems. This leads to our third conjecture: 3.2.1.1. How to achieve system-centered adaptation for produc￾ers? Recommender systems can either be fueled by explicit elici￾tation methods (ratings) or implicit methods. The positive link between participation and learning speaks in favor of explicit elic￾itation methods rather than unobtrusive and implicit methods. Through rating activities, learners are required to reflect on the merits of a recommended item, and this might function as a valu￾able metacognitive prompting strategy. 3.2.2. Social adaptation In Section 3.1.2 on the recipient role it was argued that human information processing has a social dimension and is colored by biases, preferences, and habits. A similar case can be made for the producer role, particularly in cases of explicit elicitation through ratings. Rating an item represents a social dilemma (Dawes, 1980). Such a dilemma occurs when (a) it appears rational for each individual to withhold rather than produce or share infor￾mation, and (b) it is better for the collective if every member con￾tributed rather than withheld information. This is the case for recommender systems where rating requires some effort, but the immediate benefit of rating is not evident to a user. Social dilem￾mas can lead to detrimental behaviors like social loafing and free-riding (Karau & Williams, 1993). As social loafing directly im￾pedes the quality of a recommender system, this raises the ques￾tion of how this detrimental behavior can be averted. As users are highly likely to respond to social cues, the basic idea here is to emphasize the social aspects of a recommender system. While recommender systems are peer technologies, there is no direct peer-to-peer interaction, and the community of users remains anonymous and invisible to an individual. However, by making the community more visible, powerful social psychological mech￾anisms can be evoked. Two strategies are built on these mecha￾nisms, and their impact on the quantity of ratings has been investigated empirically. The first strategy makes use of the nor￾mative power of groups, either through introduction of social com￾parisons or through goal setting. Depending on prevalent social identity (Tajfel & Turner, 1986), individuals adhere to group norms, and this adherence can even be stronger when members are anon￾ymous (Reicher, Spears, & Postmes, 1995). If the group norm is about member productivity, rating quantity can be increased by introducing social comparisons (Festinger, 1954). This was investi￾gated by Harper, Li, Chen and Konstan (2007). They gave bogus feedback to participants of the movie recommender system Movie￾Lens which indicated that the number of ratings that an individual has provided was lower, the same, or higher than a comparable group of community members, and contrasted this to a condition without such feedback. It was shown that upward comparison (feedback about under-performance) led to the highest number of produced ratings in the following week. Even downward com￾parison (feedback about over-performance) led to a higher number of ratings than the control condition without social comparison information. A different normative approach was investigated by Ling et al. (2005) who reported that setting of concrete norms and goals like rating a fixed number of items led to higher produc￾tivity than setting unspecific ‘‘do your best’’-goals, at least when the goal seemed attainable. Taken together, it appears that making norms of a community salient can exert normative power which in turn increases productivity. The second strategy to appeal to the social nature of recommender systems is by making one’s contri￾bution more valuable. The collective effort model (Karau & Williams, 1993) posits that social loafing will be reduced when people believe that their contribution is useful to a community. Moreover, it states that people contribute more when they identify with similar others. In the digital library example of Section 2, our student had a taste that differed from the mainstream. In this re￾gard, her ratings are particularly valuable for the sub-group of like-minded people. But she might neither know that her taste is special, nor that there are like-minded people. Therefore it would be helpful if some information on the utility of ratings were pro￾vided. This issue has been investigated in three recommender-re￾lated studies. Two of these studies manipulated utility by telling subjects that they either had a very unique taste (high utility) or a very typical taste (low utility) (Ling et al., 2005; Ludford, Cosley, Frankowski, & Terveen, 2004). The authors confirmed the predic￾tion of the collective effort model that uniqueness instructions lead to more contributions. A third study, conducted by Rashid et al. (2006) used a more technology-oriented approach to employ util￾ity information. The authors created a recommender interface where each unrated item had a display that indicated how helpful it would be for target persons if it were rated. In line with the col￾lective effort model, it was found that displaying the rating utility led to higher contribution rates than a control condition. It was also confirmed that people felt more motivated to contribute for the good of similar target persons than dissimilar others. Contrary to expectations, fewer items were rated when the benefit to one￾self was stressed. This is somewhat surprising, as the quality of rec￾ommendations for an individual increases with the number of ratings that this individual has produced. However Herlocker, Konstan, Terveen, & Riedl (2004) have pointed out that motivations for user ratings can be quite different: Some users simply want to express themselves; some users are driven by social motivations like helping others, or manipulating others; and for some users, gaming the system is the main motivation to provide ratings in a recommender system. In all these examples, utility for oneself does not play a major role. In contrast, making the social impact of one’s recommendation salient is an effective method to boost rating activities. This leads to our fourth conjecture. 3.2.2.1. How to achieve social adaptation for producers? This can be accomplished either by making group norms visible (through so￾cial comparisons or concrete goal setting), or by providing informa￾tion about the usefulness that rating provides for (similar) others. For educational contexts, the strategy of stressing the utility of rat￾ings might be superior to the social comparison strategy, as both upward and downward comparisons were reported to be associ￾ated with negative affect (Buunk, Collins, Taylor, van Yperen, & J. Buder, C. Schwind / Computers in Human Behavior 28 (2012) 207–216 213