Social recommender Systems: Recommendations in Support of E-Learning Sheizaf rafaeli: Yuval Dan-Gur: Miri Barak International Journal of Distance Education Technologies, Apr-Jun 2005; 3, 2; ABI/INFORM Global 30 Journal of Distance Education Technologies, 3(2), 30-47, April -June 2005 Social Recommender systems Recommendations in Support of E-Learning Sheizaf Rafaeli, University of Haifa Mt. Carmel, Israel Yuval Dan-Gur, University of Haifa Mt Carmel, Israel Miri Barak, Massachusetts Institute of Technology, USA ABSTRACT Recommendation systems can play an extensive role in online learning In such systems, learners can receive guidance in locating and ranking references, knowledge bits, test items, and so forth In recommender systems, users'ratings can be applied toward items, users, other users ratings, and, if allowed, raters of raters of items recursively. In this chapter, we describe an online learning system-QSIA-an active recommender system for Questions Sharing and Interactive Assignments, designed to enhance knowledge sharing among learners. First, we lay out some of the theoretical background for social, open-rating mechanisms in online learning systems, We discuss concepts such as social versus black-box recommendations and the advice of neighbors as opposed to that of friends. We argue that enabling subjective views and ratings of other users is an inevitable phase of social collaboration systems. We also argue that social recommendations are critical for the exploitation of the value associated with recommendation. Keywords: collaborative filtering: friends, knowledge sharing; neighbors; QSIA; recommendations INTRODUCTION communication among learners and trans fer of information. They offer opportuni E-learning involves the use of a com- ties for enhancing ways in which teachers puter or electronic device in some way to teach and learners learn(Hoffman, Wu ng its many Clark and Mayer (2003)define e-learning applications, the Web serves as a tool for as instruction delivered on a computer by way of CD-ROM, Internet, or intranet that Barak, Addir, 2003; Eylon, 2000; Rafaeli is designed to support individual learning or ravid 1997) and the creation of learn organizational performance goals. Th ing communities( Gordin, Gomez, Pea,& Internet and the World Wide Web (www) Fishman, 1997; Sudweeks Rafaeli, 1996) facilitate e-learning by allowing worldwide Ihe spectrum of knowledge items on the Copyright o 2005. Idea Group Inc. Copying or distributing in print or electronic forms without written permission of Idca Group Inc, is prohibited Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Social Recommender Systems: Recommendations in Support of E-Learning Sheizaf Rafaeli; Yuval Dan-Gur; Miri Barak International Journal of Distance Education Technologies; Apr-Jun 2005; 3, 2; ABI/INFORM Global pg. 30
Journal of Distance Education Technologies, 3 (2), 30-47, April-June 2005 31 Internet runs from useful, fascinating, and delivery of popular items, as suggested by important to pointless, bizarre, and mislead Maltz(1994). Konstan et al. (1997)found ing For learners who wish to gain knowl- personalized (rather than impersonal)pre- edge by using information and communi- diction to be significantly more accepted cation technologies(ICT), the actual ben- by users. Recommendations we receive efit of what they stand to gain will be af- daily rely mainly on human-analyzed fected by how well they make discerning sources: movie reviews, rumors, word-of- judgments about what they find( Burbules mouth, surveys, guides, friends, and rec callister, 2000) ommendation literature (Shardanand Judicious use of ICT can boost learn- Maes, 1995; Resnick Varian, 1997) ing that is adapted to the abilities of each Recommender systems approach the prob tudent and enhance the distribution of lem of helping users find preferred items knowledge among users. Psychologists mainly with the technique of Collaborative make distinctions between explicit and tacit Filtering(CF). The basic idea of CF algo- knowledge. Explicit knowledge is the rithms is to predict the likeliness list of the knowledge that can be written down, top-N recommended items based on the whereas tacit knowledge is the knowledge opinions(either explicit or implicit)of like- that lies in the learners'minds Capturing minded users(Sarwar, Karypis, et al and sharing tacit knowledge is extremely 2001); the task is to predict the utility of difficult and was the aim of various studies items for a particular user( the active user), (Kakabadse, Kouzmin, Kakabadse, based on a dataset of users' votes from a 2001). While digitized content in any form sample of population of the other users is explicit knowledge, not many e-learning However, many recommendation systems approaches encourage learners to provide produce unsatisfactory results(Herlocker, their tactic knowledge Konstan, riedl, 2000; Oard Kim, Recommendation systems can play 1998) a large role in online learning as providers Recommendations carry different of tacit knowledge. In such systems, learn- values for the provider, as contrasted with ers can receive guidance in locating and the recommendation seeker. Rafaeli and ranking references, knowledge bits, test Raban(2003)show how the Endowment items, and the like. The core task of a Effect (an extension of the Prospect recommender system is to recommend (in Theory)exists with respect to information a personalized manner)interesting and valu- as well -people value information they able items and to help users make good own much more than information not owned choices from a large number of alterna- by them. Accordingly, research shows that tives without having sufficient personal users tend to ask for recommendations experience or awareness of the alterna- more than to supply them(Avery, Resnick, tives( Gordon, Fan, Rafaeli, Wu, farag, Zeckhauser, 1999; Herlocker, Konstan, 2003: Grasso, Meunier, Thompson, 2000; riedl, 2000). Recommendation systems Oard Kim 1998: Resnick Varian, can be based on human resources--social The task recommend items to a user, algorithms-black-boxes. We argue that hen referenced in recommender systems a large portion of the shortcomings of rec- research, should be interpreted mostly by ommendation systems can be understood personalized manner and not solely by the as a failure to construct social recommen oright 2005, Idea Group Inc. Copying or distributing in print or electronic forms without written ssion of Idea Group Inc. is prohibited. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
2 Joumal of Distance Education Technologies, 3(2), 30-47, April-June 2005 dation systems as opposed to black-box items collaboration. Learners should be (non-social)ones. In this chapter, we dis- responsible for recognizing and judging pat- cuss the role of social recommendation terns of information and then organizing systems for supporting learning and com- them, while the computer system should munities of learners In the following sec- perform calculations, store, and retrieve tion, we provide a literature review about information (Jonassen, Carr, Yueh, 1998) the importance of social capital and de- Rather than using the power of computer scribe a recommendation system for knowl- technologies to disseminate information, it edge sharing in learning should be used in all subject domains as tools for engaging learners in reflective, RECOMMENDATION FOR critical thinking about the ideas they are E-LEARNING: studying (Jonassen, Carr, Yueh, 1998) EXPLOITING THE These statements justify another interest- ing e-learning mental model-the one that LEARNING NETWORK recognizes the social and contextual char. AND TACIT KNOWLEDGE acter of e-learning. From this perspective, FOR EFFECTIVE the exploitation of social capital and unexploited tacit knowledge of learners is LEARNING a critical challenge. Open, public views in general and E-learning mental model varies in dif- good recommendations in particular en fere approaches. Several esearchers when facing too anticipate e-learning as a solid technologi- many items, the ability to focus on the best cal phenomenon in which the selection of and ignore the rest is a necessity. The rec a platform can promote the desired learn- ommendation process is a social one ing outcome. In recent years, this approach recommender systems form a community has been evaluated as extremely simplis- of people voting and expressing their opin tic. Several other issues have to be consid- ions about items in a domain on one hand ered in order to expand the value proposi- and seeking recommendations on the other tion of e-learning. Even though the evalua- Being a part of such a community involves tion criteria for efficiency of e-learning sys- social dilemmas-the effort one tends to tems are not agreed upon yet (Lytras et invest in recommending an item(e. g,Am al.,2003a,2003b, 2003c), the dominant sci- I only a small, unimportant part of a large entific opinion is that technology functions group? How much does my specific opin as a mean and not a goal in the learning ion count? As to receiving a recommenda process(Lytras et al., 2003a). In addition, tion, who participates in my recommend learning is influenced by participation in a ing group? On what basis of similarity do community(Bruner, 1990; ygotsky, 1978); opinions aggregate?) learning also involves the use of many re when investigating the role technol sources. In order to sort and select the suit- ogy plays in e-learning systems, Lytras, et able resource, learners seek guidance and al. ( 2003)look at both the process of learnin commendations. The process of seeking and at the product (i.e, learning content in and providing experience-based recommen- terms of learning objects)and reach the dations across users'communities is one conclusion that care should be taken to the way of implementing large-scale knowledge appropriate balance between the role of Copyright 2005, Idea Group Inc. Copying or distributing in print or electronic forms without written permission of Idea Group Inc, is prohibited Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Journal of Distance Education Technologies, 3(2), 30-47, April -June 2005 33 technology and pedagogy in e-learning system's quality of performance (some failures of systems are explained by Recommender systems are often accom overestimation of either one as opposed to panied by some misconceptions about their a well customized solution). Modeling the function or process. Many often think that learning process and the learning product recommender systems are exhibits a clear view on the multi-dimen sional spectrum of e-learning systems At no cost, in most cases, the user can (Lytras et al., 2003a); one of its main in eceive free recommendations sights is, when referring to e-learning, A collection of a large number of items knowledge management is not a techno- many of which may not be relevant logical phenomenon; it's a qualitative shift (Grasso, Meunier,& Thompson, 2000 in people's behavior.. "(p. 2580). From Oard Kim, 1998; Resnick& Varian, this perspective, our work further extends 1997) the discussion of the content perceptions Thought to be more objective and ratio in e-learning systems. The diffused knov nal than human advisers(Dijkstra edge within e-learning systems cannot b Liebrand et al., 1998) seen as a solid one, but has to exploit fur-. Thought to be accurate, though evalua- tion of computerized advice is reported In the next section, we go a step fu to be biased(Dijkstra, 1998; Dijkstra ther. We identify a number of challenges or recommendation systems in e-learning,. Thought to be trustworthy, as comput- and we elaborate further their key idea to ward the enhancement of knowledge ex erized systems make information look ploitation in implementations. more credible(Dijkstra, 1998; Dijkstra, 1999: Dijkstra, Liebrand et al., 1998; KEY CHALLENGES OF Murphy Yetmar, 1996) RECOMMENDATION Recommendation systems research SYSTEMS IN E-LEARNING is confronted with this reality: many (if not most) recommendation systems produce a lea due to its huge knowledge repository, unsatisfactory results(Herlocker, Konstan, rning community must develop the Riedl, 2000; Oard Kim, 1998). We ability to mine relevant lea objects have previously listed sor (Lytras et al., 2003c)-the recommenda- nesses related to systems'failures(Rafaeli tion layer in an e-learning system is of great &Dan-Gur,2002): importance to this goal It is important to notice that even ob-.Black boxes: provide no transparency jective indices in the field of recommender into the working of the recommendation systems are not always agreed upon (Herlocker et al., 2000). In the e-learn (Burke, 2002; Pennock, Horvitz Giles, ing context, this fact requires a better fit 2000; Soboroff, Nichols, Pazzani, 1999) of the whole system within the e-learn- and the fact that human taste suffers from ing information systems. It is required noise(Freedman, 1998; Pescovitz, 2000) to integrate recommendations as an in doesn't make it easier to determine a evitable part of the e-learning unique experience Copyright 2005, idea Group Inc, Copying or distributing in print or clectronic forms without written permission of Idea Group Inc is prohibited Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
34 Journal of Distance Education Technologies, 3(2), 30-47, April-June 2005 Cognitive load effort: a high load is lem in e-learning systems. Learners en- required in the process of assigning ac- roll in courses again and again, and, ac- curate explicit ratings, making it difficult cording to the diffusion model of e-learn to assemble large user populations, thus ing content, they y can pre ontributing to data sparsity(Oard recommendations. The interesting issue Kim, 1998). In e-learning, this issue can for investigation refers to the quality of be addressed by incorporating the rec- the relevant recommendation ommendation process as an integral part Data sparsity and first rater problem: of learning content exploitation. The the number of people who rate items is challenge here is to design transparent relatively small, compared to the num terfaces that capture the recommen- ber of items(Terveen Hill, 2001),es dation without posing extra anxiety to pecially with regard to why anyone e learner while he or she uses the e- should volunteer to rate a new item. In e-learning, the rating of a learning item Exploration/Exploitation tradeoff: incorporated in the learning whether to recommend a wider range process. A functional way is to provide of items about which there is less cer- recommendation, such as a hidden task, tainty or only those which are known to after the completion of the relevant match the user profile learned so far learning session. In this way, the learn- Balabanovic, 1998) ing object paradigm for structuring and Initial user profile: sometimes re- exploiting learning content is quite inter- ferred to as "user model. "this is diffi- esting. A knowledge provider could be cult to form, though it has great impor- the first rater, and consequent tance(Maltz.& Ehrlich, 1995). In an e the learning content can support the learning system, learner profiles can be elaboration managed more effectively. A learner is Performance speed: systems with a a person whose prior knowledge, cogI arge dimensional number (items X us- tive level. and so forth can be outlined ers)slow down online computing per before the learning experience formance(Sarwar, Karypis et al., 2001) New item: a new item in the systems database has no ratings and cant be This list outlines a number of inter- recommended until more data is ob- esting problematic issues that have to be tained(Balabanovic Shoam, 1997). considered in e-learning. The fact is that From this perspective, recommendations e-learning context differentiates from a challenge the learning content creation business environment, since the learning process: The semantic annotation of content has to be valued in terms of learn content can be justified on a social net- ing needs. In other words, recommenda- work basis: Threads of recommenda- tions can be expressed in several different tions from several learners complete the formal metrics with an emphasis on peda value perception of learning content. gogical value, learning need, suitably for Economy of scale: a large user popu- working task, and so forth. In this way,our lation is needed to produce reliable re- work is investigating the social recommend sults( Balabanovic Shoam. 1997: Im ing systems as a key answer to the black Hars, 2001). This fact is not a prob- box approach. This discussion is presented opyright 2005, Idea Group Inc. Copying or distributing in print or electronic forms without written permission of Idea Group Inc. is prohibited Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission