Interest-Based vs. Social Person-Recommenders in Social Networking Platforms

grohg,brocco,@in.tum.de; kleemu@web.de Inter est-Based vs. Social P erson-Recommenders in Social Networking Platf orms Georg Groh 1 , Michele Brocco 1 , and Andreas Kleemann 2 1 Fakult ¨ at f ¨ ur Informatik, T echnische Univ ersit ¨ at M ¨ unchen, Germany 2 Utopia A G, M ¨ unchen, Germany Abstract. Social network based approaches to person recommendations are com- pared to interest based approaches with the help of an empirical study on a large German social networking platform. W e assess and compare the performance of different basic variants of the two approaches by precision / recall based perfor- mance with respect to reproducing kno wn friendship relations and by an empiri- cal questionnaire based study . In accordance to expectation, the results show that interest based person recommenders are able to produce more novel recommen- dations while performing less well with respect to friendship reproduction. With respect to the user’ s assessment of recommendation quality all approaches per- form comparably well, while combined social-interest-based variants are slightly ahead in performance. The overall results qualify those combined approaches as a good compromise. 1 Introduction The term Social Recommender Systems can be understood in a variety of ways. The first interpretation of the term may substitute the actors of some sub-netw ork of a social network for the set of users with similar rating-behavior as a neighborhood for making collaborativ e recommendations (see e.g. [20], [4] or [13]). This approach has been shown to possess certain advantages ov er traditional col- laborativ e filtering [15] and has been sho wn to be able to perform as good or better at least in taste related domains [13][4]. The adv antages encompass a better perfor- mance in certain situations in vie w of the portfolio ef fect (a user is recommended items which he already knows or which are too similar to those he already knows (see [2])) or cold start ef fects (with respect to ratings and trust [12]). New influences (structurally or radically new recommendations) can enter the information space of a user through sub-networks of her social network (horizon broadening effect) as easy of ev en easier than through groups of similar rating b ut otherwise unrelated users. But the external intelligence injected into the system by explicifying and using direct or indirect social relations as in social networking platforms may ensure a some what higher probability of relev ance of those radically new and unexpected, horizon broadening recommenda- tions for the user . Reasons for this can be normativ e effects in groups [6][19] (“I should know and like what my peer group likes”) or trust effects and easier explanations for recommendations via the social network (“I trust and know how to value this album recommendation based on Mark’ s, Jenny’ s and Y iming’ s musical information space be- cause I know them and their relation to me and their function, role, position etc. in the network”). There may , ho wever , be some cases where social recommenders in this first interpretation may be less useful than recommendations by “network” of anonymous but similar rating users (e.g. in case of recommendations of scientific papers) where implicit “topical” relations to these users are exploited. A second interpretation of the term Social Recommender System may encompass recommending items not to single users but to whole groups of users. In this interpre- tation the target of the recommendation (the group) is a socially defined concept [3]. A third interpretation of the term may make persons or groups of persons the rec- ommended entities, either using social filtering (as discussed above in the first interpre- tation of the term), con ventional collaborati ve filtering, content-based filtering (using any accessible electronic representation of the recommendable persons or groups as a basis for similarity computations) etc.. One example for this interpretation are team recommendation systems (see e.g. [5]) where teams are recommended (e.g. to HR ad- ministrators) especially in situations, where the number possible team configurations is very high (such as Open Inno v ation scenarios). This contribution will deal with a flav or of third interpretation: In a social network- ing platform recommending potentially interesting other users to a user based on mutual interests. While user recommendations on the basis of the social network are quite com- mon today (consider e.g. the friend recommendations in Facebook), the problem of ho w to assess and incorporate user’ s interests into the recommendation in a simple and ex- pressiv e way is still subject to research. In this article, we in vestigate the question, ho w simple interest based person recommendation approaches performs in contrast to social network based recommendation approaches. In Section 2 we revie w more related w ork concerning social recommender systems. Section 3 describes the setting of our study , the data-set we use, and the range of recom- mendation methods we inv estigate. Section 4 then presents and discusses the results. On the one hand we compute performance measures on the basis of reproduction of kno wn friendships as indicators of the usefulness of the recommenders. Furthermore the results of an empirical study among our test-users is presented and discussed. In the conclu- sion, we summarize the overall results and shortly discuss and compare the implications for the various approaches to person recommendations. 2 Related W ork Currently the most common example for people recommendation is people match- making in social networks. Popular application domains therein are dating platforms or expert finders. The first, may consider also preferences on different scales besides traditional demographical data such as age, gender, etc. The second considers skills, competencies and expertise acquired from v arious sources in order to recommend an expert that could pro vide suggestions and help for a specific problem. In online dating platforms persons usually fill in data that should describe them- selves appropriately in order to find partners that match their person in terms of de- mography (e. g. age, gender , height), interests (e. g. para-gliding, watching movies) and preferences (e. g. rock music, smoking). A few of them allows for entering a so called “target profile” that represents the description of the person one would like to receive as recommendation. Fiore et al. for instance, in vestigated which data within a profile influences the percei v ed attracti veness by women and men by correlating the percei ved attractiv eness with various elements of a user’ s profile (photos, free-text components and fixed-choice components) [11] . Diaz et al. [9] describe the match-making problem from an information retriev al perspectiv e and propose a nov el approach for the combination of user profiles to im- prov e the rele vance of recommendations. There, features are extracted from a user pro- file (e. g. free-te xt descriptions) and used as input for a machine learning algorithm that selects the most important predictors for good matches. Good matches were considered those matches where bilateral user interaction could be identified or the same features applied as in the conditions where a bilateral contact occurred (cp. labeled vs. predicted relev ance). Expert finders are a dif ferent domain for person recommendations. There primarily competencies are reg arded in order to increase the probability to find a solution for an occurred problem. In the simplest case, the task of finding experts can be solved with simple database queries. Howe v er , this does not always entail satisfactory results due to the difficulty to formulate appropriate queries and because skills may not be the only criterion for searching. McDonald and Ackerman [17] for instance tried to model current best practices for finding experts in a large company and mapped these heuristics to a corresponding system. In another work McDonald augmented this system with two different social net- works: one based on workplace sociability , which represents how often individuals so- cialize with each other, the other based on shared workplace context, which represents logical work groups and work context ov er org anizational boundaries [16]. His work emphasized that it is challenging to mix skills and social networks in recommendations because users percei ve a trade-of f between the two: more precisely e ven though the sys- tem looks first for experts and ranks them afterwards according to the social network, the users think they get only recommended due to the latter aspect. A related system described by Ehrlich et al. [10] was used (among other functionalities) to recommend experts searched by keywords within a specific social distance in a user’ s social net- work. Furthermore, this work addresses also other aspects such as pri v acy , acceptance and usage of such systems. Guy et al. [14] describe a slightly different way of recommending persons. Their approach bases on the collection of data (from blogs, social bookmarking, etc.) in a company’ s internal intranet in order to make suggestions for adding people that may belong to its social network but were not explicitly added to the social networking platform. Regarding person recommender system in enterprise-internal social networking plat- forms, Chen et al. [8] dev eloped a person recommender based on ke yword extraction algorithm, that tries to extrapolate user interest from user contributions. This approach is very valuable as foundation for our purpose, ev en if the findings of enterprise-internal platforms can not be inherently applied to comminties of interest such as the utopia community . Additionally , in contrast to Chen et al. we rather want to extrapolate user interests from all user activities performed in the predefined topic categories provided by the platform and hence inv estigate whether this kind of categorization technique is suitable as background data for interest-based person recommendations. As a last as- pect, the recommendation provided by Chen et al. ’ s system, may have a different kind of utility or goal (see Section 3) because of the business domain the community is situated in. An extensi ve revie w on social matching systems is provided by T erveen and Mc- Donald [21] that additionally formulate claims and related research questions in this field. This work can be used to derive guidelines for the de velopment of systems that incorporate social aspects (especially social networks) to find appropriate matches. 3 Data-Set and Methods W e had access to the complete database from the German based social netw orking plat- form Utopia.de [1]. The main purpose of the platform is the collaborative promotion, discussion and dev elopment of ideas and concepts contributing to more en vironmental sustainability . The platform provided the usual set of services and data-elements, like priv ate messaging, discussion boards or blogs and personal profiles, which, in contrast to platforms targeted at self presentation, are rather sparse and use few pre-determined elements. In contrast to social network based friend recommendations (which we will refer to in the following as friend of a friend (F oF) recommendations) being widespread in Social Networking platforms or content-based recommendations comparing only the profiles directly , we are interested in recommending users other users on the basis of their interests reflected by their actions and content on the platform. Thus the sparse profiles are not a problem. Besides user generated contributions, the platform also contains editorial material which can be commented upon by the users. Furthermore, users can express positi ve attitudes to ward a contribution by assigning it a “worth living” point. Instead of a free social tagging system, the platform has eleven content categories ( C 1 , . . . , C 11 ) like e.g. “Health and Diet”, “Construction and Renov ation” that users can attach to any con- tribution, which can be vie wed as a simple form of tagging with a fixed tag set. Social tagging based person recommender systems (e. g. [22][18]), in general, recommend per- sons according to similar tagging behavior . In contrast to classic Collaborativ e Filtering (CF) which basically uses similarity measures on the columns of the user-item-rating matrix R { ui } for neighborhood creation to wards recommending items, these systems use the user-tag-item matrix T { uti } to identify users with similar tagging behavior for recommending these persons (e.g. as a means of expert finding). Classic CF belongs to a class of recommending approaches that use explicit ratings of items (for item recom- mendation) or of persons (for person recommendation), whereas social tagging based approaches belong to a class of methods that use implicit methods. Implicit methods induce user attitude tow ards items or similarity to other users indirectly from their be- havior on the platform (e.g. frequency of accessing certain contributions) or the content in their information spaces (their contrib utions on the platform or their profile, which can be compared using techniques from information retrie v al (e.g. tf-idf vectors and cosine similarity)). For person recommendations, users with similar tagging behavior can be consid- ered to ha ve “similar interests” and are thus candidates for being recommended. In essence, this interpretation is not necessary . The term “users with similar interests” can be considered synonym for “users that are similar with respect to their behavior on the platform and / or their information spaces”. In contrast to having to compare users by comparing matrices as sim ( u 1 , u 2 ) = sim ( T u 1 { ti } , T u 2 { ti } ) as in social tagging based person recommenders, a simpler ap- proach is to count all platform acti vities of a user related to a certain cate gory C i (“cat- egorized activities”). Such categorized activities can be the creation or commenting of a content item (e.g. a blog entry) or the assignment of a “worth living” point. For each user u , these counts act i are then normalized with the total number of categorized ac- tivities P i act i , to yield the normalized categorized acti vities A i = act i / P i act i . W e can then compare these vectors as sim ( u 1 , u 2 ) = sim ( A u 1 { i } , A u 2 { i } ) to yield a sim- ilarity measure for users. For the actual comparison of the vectors, we use standard cosine similarity and Pearson correlation. A minimum total number of categorized ac- tivities is necessary to be included in the matrix. From the resulting similarity matrix, we recommend users with a similarity above an adjustable threshold v alue that were not already “friends”. Unfortunately , in the absence of a free social tagging mechanism in the platform we cannot compare the performance of our approach against the social tagging based person recommenders discussed before. An interesting question re garding the validation of person recommender approaches is the performance metric to be used. At this point a div ersification of the dif ferent rec- ommender goals should be done since two possible goals can be generally pursued. One possible ev aluation method is to ev aluate whether the user accepted the recommenda- tion (for instance by clicking on the recommended item). The other possible ev aluation is whether the recommendation itself is useful, i. e. if the person recommended in fact does fulfill a user’ s expectations (with respect to a predefined goal such as e. g. for a friendship, as discussion partner , as expert). Obviously , a strict diversification of goals is not possible, because (i) the goals are conceptually not completely disjunct and (ii) from a technical point of view platforms do not provide this div ersification for classi- fying users. Thus, friendships in social networks are treated as a sort of “bookmarks” to find persons with respect to all the abov e mentioned goals. In the utopia case, that can be reg arded as a community of interest, we want people to get in contact aiming at finding new discussion partners such that interesting hints and suggestions related to the topics discussed in the utopia platform can be better exchanged. Howe ver , as mentioned the target platform does not pro vide any diversification con- cerning this aspect. For this reason, and kno wing that the goals for person recommenda- tion in our case overlap, as a measure for the potential success of the approach we had to choose the reproduction rate of friendship ties that already exist in the platform as ev aluation criterion. Therefore, and also in order to compare the approach against a F oF based approach, we e xclude members with less than 3 friends and less than 8 friends of friends within the test group. The minimum total number of categorized activities was set to 3. The resulting test group encompassed 334 users with 3984 friendship relations. The mean number of friends was 11 . 93 . mean number of FoF was 270 . 31 and mean number of categorized activities was 87 . 56 . ∼ 31% of the test users had only 3 or 4 friends, and only ∼ 17% of the users had only three or four categorized acti vities. The FoF Recommender that is similar to the friend recommenders used in common Social netw orking platforms (see Section 2) and that we compare our approach against, recommends a person u 1 to a person u 2 in proportion to the number of common friends f u 1 ∧ u 2 relativ e to the average total number of friends of both users 0 . 5( f u 1 + f u 2 ) . The FoF similarity or recommendability of u 1 and u 2 is thus given by sim ( u 1 , u 2 ) = 2 f u 1 ∧ u 2 / ( f u 1 + f u 2 ) . W e used 10-fold cross v alidation in our experiments: For each of 10 runs of all of the recommender approaches that we compare, we leave out one tenth of the friend- ship relations and use the remaining nine tenth of the friendship relations to compute FoF Recommendations. The data basis for most of the variations of the basic interest based recommendation approach, which will be discussed in the next section, remains constant. W e then compute the n = 10 best recommendations for each user and each ap- proach and measure how many of the deleted one tenth friendship relations are “repro- duced” by the recommender . If we recommend a total of A persons in one run and ha ve 398 deleted friends per run, we can determine the true and false positi ves ( T P and F P ) and the false negati ves F N and hav e A = T P + F P and F N = 398 − T P . W e can then compute Precision, Recall and F-Measure as usual as measures of the success rate. If the random 10-fold partitioning of the friendship relations deletes less than 1 or more than 10 friendship relations for a single user , we do not compute recommenda- tions for this user . In these cases we cannot determine the success rate analogous to the “regular” cases. Thus from the 334 (users) ∗ 10 (runs) = 3340 cases we only compute recommendations for 1921 of these cases. Since for each case we recommend the top n = 10 best recommendations we make 19210 recommendations in total. If we recom- mend a person that is already a friend in the respective nine tenth relation data set, we drop this recommendation. W ith this procedure, we can, of course nev er reach precision values of 1, simply because we delete on a verage in each run only 2 . 07 friend relations and recommend al- most always 10 persons. Howe ver , these restrictions apply to all recommenders equally . 4 Results and Discussion A general strength of the proposed approach can be that, in contrast to social network based approaches like FoF , a user does not need a friend-list, but a weak point is that passiv e users that do not perform many explicit actions will not acquire a meaningful A { i } vector . T able 1 shows the basic results of the experiment. What we see from the table is that the interest based recommender approach is significantly better than random in re- producing pre-existing friendships. The FoF approach is ev en significantly better . This can be attributed to the fact that even in a platform that is mainly targeted towards ex- change of content in vie w of a narro wer field of interest (a typical community of interest [7]), friendship relations are perceived mainly as something social and not so much as something content or interest related. The formation of social friendship ties will ob- viously be strongly influenced by the friend of a friend effect and can thus much more easily be reproduced by the FoF recommender . Howe ver , our approach does not aim Recommender total # recommen- dations true positiv es (TP) (reproductions) precision recall f-measure Random 19210 144 0.008 0.036 0.012 Interest Based Pearson 19048 250 0.013 0.063 0.022 Interest Based Cosine 19210 283 0.015 0.072 0.024 FoF 19132 1164 0.061 0.294 0.101 Interest Based Pearson plus link 19048 376 0.020 0.095 0.033 Interest Based Cosine plus link 19210 422 0.022 0.107 0.036 T able 1. Results of the experiment. 10-fold cross validation: precision, recall and f-measure av- eraged ov er 10 runs. at recommending friends in the mere social sense, but rather at recommending persons that are related via interests in platforms where the exchange in terms of content is the main goal as opposed to platforms where self-presentation and socially related commu- nication is predominant. Of course, the social sphere and the interest based sphere are closely related. An attempt to ne vertheless impro ve our interest based approach, we in vestigated, if weighting dif ferent sorts of categorized actions differently can make a dif ference (e.g. by giving creating a long blog-entry a higher weight than just assigning some item a “worth li ving” point). The variations only very slightly improv ed the performance (e.g. in the Pearson case plus 4 % for precision), which does not allo w for any significant conclusions. In [8], authors were able to improv e a content based item recommender system by additionally taking into account social relations between the item’ s owners. In accor- dance to that we also in vestigated, in how far our interest based person recommendation approach may profit from combining it with a social relation based component. W e do this by multiplying the relev ant ( ≥ 0 . 5 ) interest based similarity scores between two users with a factor of 1 . 5 if the two users have at least one common friend. Thus we effecti vely augment the interest based approach with the FoF approach (and not vice versa). Thus the general advantages of our approach discussed before can be main- tained. W e called these variations “plus link” and the results are also shown in table 1. As expected, we see that the approach can profit from this augmentation by approxi- mately 50 % increase of performance. Howe ver , it has to be stressed that, as discussed before, the performance with respect to reproducing existing friendship ties is certainly not our main goal and not the only quality criterion of an interest based person recom- mender in our sense. Q-nr Question Scale 1 Do Y ou kno w this user already? yes / no 2 Are Y ou interested in getting to kno w this user? 5 point Lickert 3 Space for comments on the recommendations Free text field 4 Are Y ou generally interested in getting to kno w ne w users on Utopia? 5 point Lickert 5 W ould Y ou like to be recommended users on Utopia? 5 point Lickert 6 Space for general comments Free text field T able 2. Online survey: questions 4.1 Empirical study In order to address this issue, we conducted an online empirical study among our test users, dividing them into three groups and providing them with 5 person recommenda- tions using one sort of recommender in each group (Interest based cosine, interest based cosine plus link and FoF). The users were asked to ev aluate the recommendations ac- cording to sev eral criteria (see table 2). General statistics with respect to this surv ey are shown in table 3. Recommender Completed ques- tionnaires Percent completed Cosine 28 28.3 % Cosine plus link 35 35.4 % FoF 36 36.4 % T able 3. Online survey: general statistics The recommendations were provided in the form of picture and username of the recommended person as specified in the platform. By clicking on either username or picture, the profile page of the corresponding person could be inspected in order to identify possible interesting characteristics of that person. Based on this knowledge the user can decide whether the recommendation proposed is appropriate or not. Recommender rec. person kno wn rec. person un- known # of rec. Cosine 40,7% (57) 59,3% (83) 140 Cos. plus Link 41,1% (72) 58,9% (103) 175 FoF 57,8% (104) 42,2% (76) 180 Overall 47,1% (233) 52,9% (262) 495 T able 4. Results of question 1 T able 4 shows that the FoF variant is more likely to recommend already kno wn persons which is socially plausible. The overall high number of recommendations of already familiar persons can be explained by the fact that due to the selection scheme of the 334 users (see pre vious section), already very acti ve users were selected that hav e a high probability of knowing each other . Rec. Quest. 1 Question 2 1 2 3 4 5 Cos. (unknown) 24.1% 16.9% 34.9% 16.9% 7.2% (known) 1.8% 3.5% 54.4% 26.3% 14.0% Overall 15.0% 11.4% 42.9% 20.7% 10.0% Cos. pl. lnk. (unknown) 5.8% 27.2% 28.2% 27.2% 11.7% (known) 22.2% 6.9% 23.6% 26.4% 20.8% Overall 12.6% 18.9% 26.3% 26.9% 15.4% FoF (unknown) 9.2% 19.7% 40.8% 14.5% 15.8% (known) 15.4% 1.9% 33.7% 26.9% 22.1% Overall 12.8% 9.4% 36.7% 21.7% 19.4% Overall (unknown) 12.6% 21.8% 34.0% 20.2% 11.5% (known) 14.2% 3.9% 35.6% 26.6% 19.7% Overall 13.3% 13.3% 34.7% 23.2% 15.4% T able 5. Cross-table question 1 (Familiarity) and question 2 (Interestingness) T able 5 shows the relation between previous familiarity and the rating of interest- ingness. W e see that the error of central tendency is present throughout the results of question 2. It is overall slightly more present for the recommender that does not make use of the social network (cosine). Ho we ver , for the recommenders that make use of the social network (FoF and cosine plus link) this tendency is slightly more prominent for the unfamiliar recommended persons than for the familiar , while for the recommender that is purely interest bases (cosine) this slight effect is re versed. As an explanation, knowing a person may make it easier to come to an expressi v e estimation apart from the less meaningful middle rating. Ho we ver , it also has to be taken into account that a main v alue for a recommender is to recommend new entities (persons in our case), where the use of these nov el recommendations often can only be properly assessed a posteriori. The results of the general questions of the questionnaire are shown in tables 7 and 6. For question 4, we see that, according to expectation, the tendency to be interested in getting to know new people on a social networking platform is quite high. There are no significant dif ferences among the three test-groups. W ith respect to question 5, we see that the recommendation service is regarded as overall positiv e but judged more critically (23.3 % negativ e (rating 1 or 2) answers in question 5) compared to the general predisposition to be interested in getting to know new people (7.1 % negativ e answers in question 4). Howe ver , the share of positive answers (rating 4 or 5) among the group which were confrontend with the recommendations from the merely interest Recommender 1 2 3 4 5 Cos. 21.4% 3.6% 35.7% 21.4% 17.9% Cos. pl. lnk. 20.0% 2.9% 17.1% 45.7% 14.3% FoF 11.1% 11.1% 16.7% 41.7% 19.4% Overall 17.2% 6.1% 22.2% 37.4% 17.2% T able 6. Results of Question 5 (General interest in person recommendation service) Recommender 1 2 3 4 5 Cosine 0.0% 3.6% 32.1% 35.7% 28.6% Cos.-plus-Link 2.9% 8.6% 28.6% 31.4% 28.6% FoF 2.8% 2.8% 27.8% 30.6% 36.1% Overall 2.0% 5.1% 29.3% 32.3% 31.3% T able 7. Results of Question 4 (General interest in getting to know ne w people) based recommender (cosine) is significantly lower (39.3 %) than for the groups that were confronted with recommendations that included the social network (60.0 % and 61.1 %). This can be seen as a hint that the social network plays an important role for people recommendations. Ho we ver , the ef fect that the use of novel recommendations often can only be properly assessed a posteriori needs to be taken into account here as well, because according to table 4, the cosine recommender proposes more nov el recommendations than the FoF recommender . The cosine plus link recommender that results in roughly the same share of novel recommendations as the cosine recommender appears to be a good compromise in view of this phenomenon. 5 Conclusion From our study it can be concluded that in social networking platforms, person rec- ommenders are services that hav e some potential to deli ver an added value for a large number of users. Purely interest based recommenders may produce more novel recom- mendations than purely social network based recommenders. With respect to the survey rating of test-users, the purely interest based approaches perform slightly worse than the purely social network based approaches. The over -proportionally good performance of the FoF approach in reproducing kno wn friendships can be attrib uted to social effects and does not hav e to be taken as a definiti ve quality criterion. Mixed approaches yield many nov el recommendation while (with respect to user rating) perform as good or ev en slightly better than the purely social network based approach. Combined social network based and interest based approaches may thus be a good compromise and a promising field of future research. References 1. Utopia. http://www.utopia.de , (URL, Oct 2010) 2. Billsus, D., Pazzani, M.J.: Learning collaborative information filters. In: Proc. 15th Int’l Conf. on Machine Learning. pp. 46–54 (1998) 3. Birnkammerer , S., W oerndl, W ., Groh, G.: Recommending for groups in decentralized col- laborativ e filtering. T ech.-Report TUMI0927, TU Muenchen (2009) 4. Bonhard, P ., Sasse, M.A.: ’kno wing me, kno wing you’ – using profiles and social networking to improv e recommender systems. BT T echnology Journal 24(3), 84–98 (2006) 5. Brocco, M., Groh, G.: A meta model for team recommendations in open innov ation net- works. Proc. A CM RecSys 09 (2009) 6. Caplan, M.F ., Miller , C.E.: Group decision making and normative versus informational influ- ence effects of type of issue and assigned decision rule. Pers. Soc. Psychol. Bull pp. 306–313 (1987) 7. Carotenuto, L., et al.: Communityspace: T oward flexible support for voluntary knowledge communities. Proc. W orkshop on W orkspace Models for Collaboration, London, UK, April 1999 (1999) 8. Chen, J., Geyer , W ., Dugan, C., Muller, M., Guy , I.: Make new friends, but keep the old - recommending people on social networking sites. Proc. Int’l Conf on Human Factors in Computing Systems (2009) 9. Diaz, F ., Metzler , D., Amer-Y ahia, S.: Relev ance and ranking in online dating systems. In: Proc. SIGIR10. pp. 66–73. A CM, Ne w Y ork, NY , USA (2010) 10. Ehrlich, K., Lin, C.Y ., Griffiths-Fisher , V .: Searching for e xperts in the enterprise: combining text and social network analysis. In: Proc. GR OUP07. pp. 117–126 (2007) 11. Fiore, A.T ., T aylor , L.S., Mendelsohn, G., Hearst, M.: Assessing attractiv eness in online dating profiles. In: CHI ’08: Proceeding of the twenty-sixth annual SIGCHI conference on Human factors in computing systems. pp. 797–806. A CM, New Y ork, NY , USA (2008) 12. Golbeck, J.A.: Computing and applying trust in web-based social networks. Ph.D. thesis, College P ark, MD, USA (2005), chair-James Hendler 13. Groh, G., Ehmig, C.: Recommendations in T aste Related Domains: Collaborati ve Filtering vs. Social Filtering. Proc. Group07, Sunibel Island, USA, Nov 2007 (2007) 14. Guy , I., Jaco vi, M., Shahar, E., Meshulam, N., Soroka, V ., Farrell, S.: Harvesting with sonar: the v alue of aggre gating social network information. In: Proc. CHI08. pp. 1017–1026 (2008) 15. Herlocker , J.L., Konstan, J.A., Riedl, J.T .: An empirical analysis of design choices in neighborhood-based collaborativ e filtering systems. Information Retrie v al 5, 287–310 (2002) 16. McDonald, D.W .: Recommending collaboration with social networks: a comparati ve ev alu- ation. In: Proc CHI03. pp. 593–600 (2003) 17. McDonald, D.W ., Ack erman, M.S.: Expertise recommender: a flexible recommendation sys- tem and architecture. In: Proc. CSCW00. pp. 231–240 (2000) 18. Nakamoto, R., Nakajima, S., Miyazaki, J., Uemura, S.: T ag-based contextual collaborativ e filtering. IAENG International Journal of Computer Science (2008) 19. Noelle-Neumann, E.: The theory of public opinion: the concept of the spiral of silence. J. A. Anderson (Ed.): Communication Y earbook 14; Newb ury Park, CA: Sage pp. 256–287 (1991) 20. Sinha, R., Swearingen, K.: Comparing recommendations made by online systems and friends. In Delos-NSF W orkshop on ”Personalisation and Recommender Systems in Digital Libraries” (2001), http://www.rashmisinha.com/articles/Recommenders_ Delos01.PDF(URLJun2007) 21. T erveen, L., McDonald, D.W .: Social matching: A frame work and research agenda. A CM T rans. Comput.-Hum. Interact. 12(3), 401–434 (2005) 22. Zanardi, V ., Capra, L.: Social ranking: Uncovering relev ant content using tag-based recom- mender systems. Proc. A CM RecSys 2008 (2008)