Personalized Recommendation of Research Papers by Fusing Recommendations from Explicit and Implicit Social Networks Shaikhah Alotaibi Julita Vassileva University of Saskatchewan University of Saskatchewan 181 Thorvaldson Building 178 Thorvaldson Building 110 Science Place 110 Science Place Saskatoon, SK, Canada Saskatoon, SK, Canada Shaikhah.otaibi@usask.ca jiv@cs.usask.ca ABSTRACT CiteULike, Mendeley), which researchers often use to manage their Combining social network information with collaborative filtering digital paper repositories and bookmark libraries, users can be recommendation algorithms has helped to alleviate some drawbacks connected through different social relations. By knowing that two of collaborative filtering, for example, the cold start problem, and has users are connected, one can infer that they possibly share interests increased the accuracy of recommendations. However, the user and therefore recommend items from other connected users. A social coverage of recommendation for social-based recommendation is low bookmarking service provides many clues for interest similarities as there is often insufficient data about explicit social relationships between users based on their behavior in the system and their among users. In this paper, we fuse recommendation that uses publication authorship. Surprisingly, however, none of the popular explicit social relations (friends and co-authors) with social bookmarking tools have used the wealth of social data they recommendations that use implicit social relations aiming to increase store to build a social RS. However, there are some studies that the user coverage with minimum recommendation accuracy loss. We incorporate social information into CF techniques to increase the found that fusing recommendations from friends with recommendation accuracy. Although such social recommenders recommendations using implicit social networks increases both perform well, the social information about users that they require is accuracy and user recommendation coverage while fusing not readily available for all users. Thus, these social recommenders recommendation from co-authors increase the coverage. have lower user coverage [18]. User coverage is the ratio of users who receive nonempty recommended sets to the total number of CCS Concepts users [3]. Previous studies also showed that there is a tradeoff • Information system applications ➝ Collaborative and social between the recommendation accuracy and the coverage of the computing, Data mining • World Wide Web ➝ Web searching recommendation [10]. For this reason, in this paper we investigate and information discovery, Web applications the user utility expressed in recommendation accuracy and coverage when different types of social relations are used. We hypothesize that Keywords using social data from two social networks where the first has higher Social network; implicit social network; hybrid recommendation; recommendation accuracy and the other has higher recommendation paper recommendation; social bookmarking websites; collaborative coverage will result in recommendation that achieve a balance filtering. between both benefits. Most of the previous studies fused different recommendation algorithms using the same data source. However, in 1. INTRODUCTION our study, we fuse different recommendations that use different types Scholarly papers both help to update researchers on new research in of social relations: explicit social relations which are the relations their areas of interest and serve as a directory of other researchers initiated by one or both users involved in the social relation, and with similar interests with whom researchers can collaborate. implicit social relations which are computed by machine using data However, as publishers, online journals, and conferences proliferate, about the user behavior in the social bookmarking tools (e.g. co- the number of new published papers has become overwhelming. For bookmarking the same paper). The social networks that we use in this this reason, many recommender systems (RSs) have been proposed to study are: two explicit social networks based on co-authorship and help readers by suggesting a list of potential papers of interest. The friendship, and three different implicit social networks based on two main algorithms used by RSs are content-based filtering (CBF) readership, co-readership and tag-based social network. and collaborative filtering (CF). CBF is based on information Section 2 discusses related work. Section 3 describes the social retrieval techniques that compare a paper’s features (e.g., title, networks used in this paper. Section 4 describes the recommender abstract, keywords, publication year) with the researchers’ features approaches used. Section 5 explains the experiments, dataset and (e.g., interests or previous search queries) to find matches [2]. In result analysis. Finally, section 6 discusses our conclusions and contrast, CF (e.g., [14]) uses the similarity of paper ratings to find future work. users similar to the target user and recommend papers that these users have liked. Hybrid recommending approaches (e.g., [25]) use a 2. RELATED WORK combination of the CBF and CF approaches to alleviate the Although the first social recommendation approach appeared as early drawbacks of both approaches. as 1997 [11], no agreed-upon definition for social recommendation Another way to overcome one or more of the CF drawbacks (e.g. existed until 2013, when Tang et al. defined social recommendation cold start problem or data sparsity) is to exploit the social ties to be any recommender system that includes social relations as an between users in the recommendations. With the advent of social extra input [24]. Thus, social recommenders are hybrid recommender networks in applications such as social bookmarking systems (e.g., systems that combine social relationships (e.g. membership, friendship, following relations, trust relations) with another familiar and similar to other users (i.e., a combination of both previous recommendation method, most commonly CF. Rather than using social networks). They showed that the recommendation from users’ only the user–item matrix as in traditional CF, a social friends outperformed the recommendation of the implicit social recommendation mechanism uses two matrices: a user–item matrix, network. They explained their result with the fact that the which represents the items that are rated by the user, and a user–user recommendations are explained to the users, who can see the picture matrix, which represents the social relations between users. Many of the contact who sent the recommendation. However, this result studies demonstrate that using social information in the could be explained also with the fact that all users belong to the same recommendation process reduces the effect of the data sparsity and community, which, in this case, was the IBM Corporation, so they all have similar interests and also know each other. cold start problems [13] and enhances prediction accuracy [20]. A wise social network recommender system (WSNRS) was proposed There are many approaches combining CF recommender with a social by Mican et al. [15]. It considered explicit and implicit social relations network based on explicit social relations between users (e.g. (e.g., implicit relations based on number of clicks to see other user’s [3,8,20,26]) for example, following on Twitter or Instagram, being profile). First, the algorithm considered the user’s connections made friends on Facebook, or in general connection that is made with the up of users who have explicit social relations with the target user. It awareness or agreement of both users. For instance, Liu and Lee [20] then considered the interactions between the target user and other compared four algorithms to recommend skin products: nearest users as well as the interactions between the target user and the neighborhood CF, social CF, a combination of nearest neighborhood webpages to calculate a trust value. If the trust value was above CF and social CF, and nearest neighborhood CF with an amplification average, the target user is an implicit follower of the other user. The of data from social friends. Groh and Ehmig [8] considered the user’s recently published resources and the favorably rated resources from friends to form the user’s neighborhood to recommend local clubs the target user’s connections are then recommended to the target user. using social CF. Yuan et al. [26] tested the effect of two explicit social Mican et al. explained and demonstrated this using a case study that networks, membership and friendship, fused with conventional CF was neither evaluated by any evaluation metrics nor compared with recommendation methods to recommend music. Bellogin, Cantador any baseline recommendation methods. Thus, the effectiveness of the and Castells [3] tested different recommendation approaches to proposed method is unclear. recommend music items using tags and social network information. Generally, very few studies incorporate social relations in the domain All previous studies in social recommendations using explicit relations of research paper recommendations. One of the existing works, between users are in the taste domains (i.e. recommending music, PubRec, is an RS that suggests to the target user, for a particular movies or clubs or restaurants. However, it is difficult to use the paper of interest, the most related papers from the libraries of other recommendation algorithms developed for a taste domain to users to whom that user is socially connected [22]. PReSA [23] takes recommend research papers because in taste domains, the number of advantage of the available data on social bookmarking websites (e.g., ratings for each item is larger than the number of ratings received per CiteULike), such as bookmarked papers, metadata, and users’ research paper. Generally, researchers seem reluctant to rate research connections, to recommend papers from the users’ connections’ papers in bookmarking systems, and there is a lack of explicit ratings libraries that are similar and popular among the users’ social in the domain of research paper recommendation. Thus, most of the connections. Both PubRec and PReSA consider the explicit research done in this area is based on citation networks and implicit relationships among users in the recommendation process. Lee and feedback about the papers, generated from user actions such as Brusilovsky studied three explicit social networks—watching tagging, downloading, or bookmarking. networks [16], group membership [17], and collaboration networks Existing research has explored also the use of implicit networks in [18]—to find the extent of interest similarities between users social recommender systems. Implicit social networks are constructed involved in those networks and compare the recommendations by inferring relationships between users that may not exist in the real watching networks produced to the recommendations traditional CF world, and the users may be unaware of them. For example, the users produced [16]. Their results showed that the watching network that belong to the same neighborhood in a CF could be considered as cannot compete with CF, that the similarities between users’ libraries part of an implicit network constructed by relating uses who gave in group membership networks are insignificant [17], and that the similar ratings to the same items. These implicit relationships have similarity between two users connected using co-authorship networks been often called “trust” [1, 7, 13, 21]. For example, in [1], a trust- is comparable to user connections using explicit networks, which aware RS is proposed that uses trust metrics to personalize the require agreement between the parties [18]. recommendations for secure skiing routes by showing information from only users the target user trusts. The trust in Moleskiing is used All the studies that have been done in the area of exploiting social to alleviate the data sparsity problem using trust propagation to infer relations in recommending research papers are based on explicit the trust values for unknown users. The FilmTrust social Web site social networks, which have low coverage. In the next section, we system proposed by Golbeck [7] recommends movies using the trust propose three social networks where the social relations between developed between users based on similar movie ratings. A study users are inferred based on their publications and their bookmarked done by O’Donovan and Smyth [21] incorporates implicit trust values papers. We test the recommendation accuracy using these social inferred from user ratings into standard CF. Massa and Avesani [13] networks as sources of information and we also test the user propose a trust graph–based RS that uses trust values given by users in coverage. Then, we compare the results with hybrid approaches that addition to similarity measures to reduce the data sparseness that fuse recommendation from implicit and explicit social networks to affects new users. The results of their experiments, performed on the provide a good balance between recommendation accuracy and user Epinions dataset, show that trust-aware RSs outperform CF. coverage. Some studies compared the recommendations produced by explicit social networks with those produced by implicit social networks. For 3. EVALUATED SOCIAL NETWORKS example, Guy et al. [9] compared recommendation produced by data Three implicit social networks (ISNs) based on users’ bookmarking from users’ friends with recommendation produced using implicit behavior are proposed as candidates to use for recommending social relations among users based on their behaviors, such as using research papers. In addition, two commonly used explicit social the same tag or co-bookmarking the same webpage. Then, they compared the results with the recommendation from people who are networks are introduced: co-authorship and friendship. In this section could have different variations, such as “socialize”, “socialization” or we describe all these five social networks. “socializing”). The relations in this network also have strengths. The strength of the relation between two users is measured by the number 3.1 Implicit Social Networks of tags they share. The assumption is that the more tags two users We consider three different social networks that connect users to share, the stronger the relationship is between them. each other based on their bookmarking and tagging behavior in social bookmarking tool. First, the readership ISN connects users to the 3.2 Explicit Social Networks authors of the papers that they have bookmarked. We assume that if We consider two social networks where the relations between users users bookmark specific papers, interest overlap exists between the are explicitly defined: co-authorship SN and friendship SN. The co- bookmarkers and the authors of the papers; this overlap increases authorship relations between two users emerge when they collaborate with the increase in the number of papers users bookmark from the in writing and publishing a research paper(s). When two users same author. The relation could be unidirectional or reciprocal. The collaborate in publishing papers, this means that they share similar relation is unidirectional if only one of the users in this relation has interest and have strong relationship. The co-authorship SN also has bookmarked the other user’s publications. The relation is reciprocal a strength represented by how many papers the pair of users has co- if both users have bookmarked each other’s publications. Figure 1 authored. The other explicit SN is the friendship SN, which has shows the relations in this network, which are depicted as black undirected relations. The relation emerges between two users when arrows. For example, the relation between user 3 and user 5 is one user invites the second user to add her to her connection (i.e. reciprocal, while the relation between user 3 and user 1 is contact) list, and the second user accepts the invitation. unidirectional; user 3 is the paper’s bookmarker and user 1 is the paper’s author. The numbers on the arrows represent the strength of 4. EVALUATED RECOMMENDING the relations. For example, the strength of the relation between user 3 APPROACHES and user 1 is five, which means that user 3’s library contains five To determine the effectiveness of the three introduced ISNs as bookmarked papers authored by user 1. sources for recommendations, we compared three existing commonly used social recommendation approaches: social recommender, combined recommender and amplified recommender. These recommendation approaches were applied previously to datasets that have explicit social relations and numeric ratings of items (i.e. rating of items using Likert scale). We applied the same approaches to a dataset that has implicit social relations and unary ratings of items (i.e. existence of the paper in the user’s library). 4.1 Social Recommender The social recommender was proposed by [3]. It simply replaces the Figure 1. Sample of relations in implicit networks anonymous nearest neighbors in the user-based CF with the target Second, the co-readership ISN connects users who bookmark (and user’s social friends in the social network. To apply the social presumably read) papers written by the same authors. If user 1 and recommender to the proposed ISNs, we found the social friends of user 2 have both bookmarked papers written by user 3, then user 1 each user and used the data from those friends in the same way that and user 2 are connected using the co-readership ISN. This network anonymous peers in CF are used - by picking the top N peers and structure is useful for users who do not yet have publications and using their bookmarked papers to find candidate papers to therefore cannot have relations in network 1. The assumption is that recommend to the user. However, in the social recommender we users who bookmark the same paper(s) also have similar interests. replaced the similarity between users that is used in the predication of The strength of the relationship is measured by the number of authors the target user’s rating for unseen items with the weighted strength whose libraries overlap. Figure 1 shows an example of the between users Ui and Uj 𝑊𝑆𝑡𝑟𝑈𝑖,𝑗 calculated as: relationships in this network in blue. For example, user 5 and user 6 𝑆𝑡𝑟𝑈𝑖,𝑗 are connected because they both bookmarked papers written by the 𝑊𝑆𝑡𝑟𝑈𝑖,𝑗 = same authors; the number of overlapping author names here is five. 𝑇𝑜𝑡𝑎𝑙𝑆𝑡𝑟𝑈𝑖 We show only a part of the graph, and it includes only one of those Where 𝑆𝑡𝑟𝑈𝑖,𝑗 is the strength of the relation between Ui and Uj and five authors (user 4). 𝑇𝑜𝑡𝑎𝑙𝑆𝑡𝑟𝑈𝑖 is the sum of all strength values between Ui and all of Third, the tag-based ISN connects users if they use the same tags to other users who are connected to her. annotate their bookmarked papers. However, we do not check whether users use the same tags to annotate the same papers. We 4.2 Combined Recommender consider the tag similarity between the entire tag cloud associated with each user. We assume that the more similar tags the users have, The combined recommender integrates neighbors from conventional the higher the interest similarity. While the previous two networks user-based CF and the target user’s social friends to construct a new are based on the papers’ metadata, this network is based on user- nearest neighborhood set for the target user [3]. We then used the generated data. To build this network, the tags used to annotate the data from users in the new combined neighbors in the papers are aggregated for each user. The data is preprocessed to make recommendation following the same way as in the social the tags comparable. We follow the method described in [19] to recommender. preprocess the tags. All tags are preprocessed by converting them to lowercase, removing the stop words, and then using the porter 4.3 Amplified Recommender stemmer tool to remove any additional letters added to the root word The amplified recommender amplifies the social friends’ preferences to eliminate the effect of the word variation (e.g., the word “social” in CF nearest neighbors [20]. First, the nearest neighborhood peers were identified by CF top-N technique. Then, if the user’s social Co-authorship SN friends were also in the top-N neighbors, we used an amplifying approach to give the preferences from those social friends more Number of co-authors 247 weight in the recommendation process. The amplifying function that Total/average number of social relations/ per user 167/1.27 we used is the one used in [20], which is given by: 𝑁𝑈 𝑈 Total/average number of the co-authors’ 4181/16.93 ) ,1)  𝑖 𝑗 publications/per user Min (𝑆𝑈𝑖𝑈𝑗 × (1 + 𝑁𝑎𝑙𝑙,𝑈 𝑗 Total/average number of their bookmarks/per 43134/174.63 where Uj is the target user, Ui is one of the Uj’s social friends, 𝑆𝑈𝑖 𝑈𝑗 is user the similarity between Ui and Uj which is calculated by CF approach Friendship SN using the papers that are co-bookmarked by both users, 𝑁𝑈𝑖𝑈𝑗 is the number of interaction between the target Uj and the user’s social Number of users who have friends 2375 friend Ui, and 𝑁𝑎𝑙𝑙,𝑈𝑗 is the total number of interactions between the Total/average number of bookmarks/per friend 360,715/99.15 target Uj and all of the user’s social friends. Because the similarity value cannot be greater than 1, we chose a minimum value between 1 Total/average number of friends/per user 6171/0.31 and the amplified value. The interactions between the target user and one of the user’s social friends were based on the type of ISN on 5.2 Metrics of Evaluating Different which we were trying to apply the approach. For example, if we use Recommenders the co-readership ISN, the number of interactions equals the number We compute several metrics from the Informational Retrieval field to of authors that both users have in common (i.e., the number of measure the prediction accuracy of recommenders. Since the data authors one or more of whose papers both users bookmark). that we have is bookmarking data, which is considered as unary 5. EXPERIMENTS AND DATASET rating (i.e. presence of absence of rating), the best metrics are precision and recall at top N. It is always assumed that the items with 5.1 Dataset higher ranks in the recommended list of items are more important We collected the data for this study from the CiteULike.org social than items with lower ranks. We calculate the precision and recall for bookmarking website. This site has been in active use since three ranks: top two, top five, and top ten. Then we compare the November 2004; it currently has 8,217,384 bookmarked papers. It results among these ranks. Precision@N (reported as P@N in Figure allows social features such as connecting users, watching users 2) and Recall@N (reported as R@N in Figure 2) are calculated with (similar to following on Twitter), and sharing references. Users of respect to the rank. For example, if Precision@10 is used, we CiteULike can import scientific reference data from other resources calculate the ratio of the true recommended items to the top 10 such as PubMed and can assign tags to the bookmarked references recommended items, and the Recall@10 is the ratio of the number of for future easy access. Using the snowball method, we crawled the true recommended items in the top 10 recommended items to the test CiteULike website, starting with 500 randomly chosen, recently set. In all of our experiments, we used fivefold cross validation active users whose publications and bookmark data we collected. approach where 20% of the user’s bookmarks are used as testing data Then, we branched to collect the users’ data for the users who had and 80% are used as training data. This process is repeated five bookmarked their publications or who had bookmarked the same times, each time with different test and training sets. Then the papers as the initial users. Table 1 shows the descriptive statistics for accuracy of the prediction is calculated. the dataset collected between December 2014 and February 2015. When N, the number of recommended items, increases, a trade-off Table 1. Descriptive Analysis of the Dataset between precision and recall metrics is observed. When N increases, the precision value starts to decrease, while the recall starts to Number of users 13,189 increase. To reduce the effect of the change of the precision and Total number of distinct papers 1,043,675 recall by increasing the N value, the F1-measure (reported as F1@N in Figure 2) is used to produce evaluation results that are more Total number of publications/bookmarks/tags 19,774/1,323,065/ universally comparable. F1 can be calculated using the following equation: 3,086,565 2.𝑃@𝑁.𝑅@𝑁 Average number of publications/bookmarks/tags 1.52/98.79/3.81 F1@N= 𝑃@𝑁+𝑅@𝑁 per user 5.3 Experiments and Results Number of unidirectional relations in readership 9,248/4,909 ISN/number of users having unidirectional In this section, the conducted experiments are described, and the relations results for each experiment are explained. Number of reciprocal relations in readership 141/209 5.3.1 Finding the Best Settings for Each ISN ISN/number of users having reciprocal relations First, we run the different recommenders described in section 4 for each of the implicit social networks (readership ISN, co-readership Number of relations in co-readership 260,361/11,484 ISN and tag-based ISN) as well as for the two explicit social ISN/number of users in this network networks (co-authorship SN and friendship SN). To find the best Number of relations in tag-based ISN/ number of 223,405/11,283 settings for each network, different neighborhood sizes (k value) are users in this network tested and different ranked lists are produced (top two, top five, and top ten). We found that the best performing settings for each network with networks. However, we used weight 1 for the recommendation if the respect to more metrics are: user has relations in only one of social networks. For instance, if we aim to fuse the recommendations produced by co-authorship explicit  Readership ISN (reciprocal relations): social recommender network with recommendation produced by co-readership ISN, but with K=5 the user has only relations in co-readership, we use the weight 1 for  Readership ISN (unidirectional relations): combined the recommendation produced by co-readership and completely recommender K= 20 ignore the co-authorship ISN for this specific user. We used this  Co-readership ISN: amplified recommender, K= 20 approach to make the recommendations more personalized. The best  Tag-based ISN: amplified recommender, K= 20 weight combination for each hybrid approach is shown in Table 2.  Co-authorship SN: amplified recommender, K= all When recommendations using co-authorship network are fused with relations recommendations from reciprocal readership ISN, the maximum  Friendship SN: amplified recommender, K= all relations accuracy is reached when the recommendations from the co- authorship network are given high weight, 0.8. However, when co- Therefore, we used these settings in the next experiment when fusing authorship network recommendation is fused with other ISNs, the data from different social networks. With compatible results with the best accuracy achieved when the weight of co-authorship was 0.3 in study in [4], small neighborhood size provided the best accuracy the case of unidirectional readership ISN and 0.1 in the case of co- results. In addition, as noted for the explicit social networks readership ISN and tag-based ISN. This is because there is a high (friendship and co-authorship), the best results were achieved by overlap between the co-authorship social relations and the reciprocal using all the of users’ social relations. This is because each user has readership relations; 58.68 percent of the relations in the co- very few social relations; the average number of relations in authorship network was discovered by the reciprocal relations in friendship networks and co-authorship networks are 0.3 and 1.27 readership network. The effect of the co-readership is less visible in respectively. the other networks, and that might be because there is a huge gap 5.3.2 Hybrid Recommendation of Explicit and Implicit between the small number of relations in co-authorship network and the large number of relations in the other networks. Social Networks Then, we used a weighted hybrid recommender to combine the When recommendations from friendship network are fused with results of recommending research papers using data from explicit and recommendations from implicit social networks, we can notice that implicit social networks. Even though there are many hybrid the maximum accuracy of the recommendations occurs when the approaches [6], we prefer to use the weighted hybrid approach weight of the friendship network is higher than the weight of implicit because it brings together all the capabilities of the combined networks. approaches in a straightforward and easy to perform way. It is a linear combination that aggregates the prediction score from different Table 2: The optimum weights for each hybrid approach (ISN recommendation approaches using a different weight for each weight, explicit SN weight) recommendation approach. The hybrid recommendation is calculated Co-authorship Friendship from the linear combination of different recommendations using Readership ISN (Reciprocal) (0.2,0.8) (0.3,0.7) 𝑊𝑟𝑒𝑐𝑖 = ∑ (𝑊𝑟𝑒𝑐𝑖,𝑆 . 𝑊𝑆𝐽 ) Readership ISN (unidirectional) (0.7,0.3) (0.3,0.7) 𝐽 𝑆𝑗 ∈𝑆 Co-readership ISN (0.9,0.1) (0.1,0.9) Tag-based ISN (0.9,0.1) (0.1,0.9) where 𝑊𝑆𝐽 is the weight for the recommender Sj, and its value ranging The results of the best weight combinations are used in the next from 0.1 to 0.9, and the sum of all weights is equal to 1. Since the experiment described in the next subsection 5.3.3. optimum weight is usually derived by examining the performance of all possible combinations [6], we used all the combinations from 0.1 5.3.3 Comparison between Recommendations from to 0.9 by gradually increasing the weight of the first recommender by increments of 0.1. We first tested the hybrid approach of the co- Different ISNs with Friendship or Co-authorship SNs We conducted an experiment to compare the recommendation using authorship network (explicit) with every implicit social network, then only ISN data, with the hybrid approach that combine we tested the friendship network (explicit) with every implicit social recommendations from ISNs with each of the two explicit networks – network. However, we used a modified version of the weighted sum the co-authorship network, or the friendship network, respectively. approach called cross-source hybrid [5]. Cross-source hybrid The results shown in Figure 2 reveal that the best prediction accuracy approach favors items that are recommended by both approaches. We is achieved when the recommendation from the friendship network is agree that items that are recommended by implicit social network fused with recommendation from ISN; this is true for all implicit recommender and explicit social network recommender are more social networks and for all measures (precision, recall and F1 at top important than items that are recommended by only one ten). However, the co-readership ISN did not help in increasing the recommender. Therefore, the above equation for weighted sum prediction accuracy. In most of the cases, the prediction accuracy hybrid approach is modified as follows: stayed the same. In the readership ISN, the prediction accuracy slightly decreased when recommendation from co-authorship SN is 𝑊𝑟𝑒𝑐𝑖 = ∑ (𝑊𝑟𝑒𝑐𝑖,𝑆 . 𝑊𝑆𝐽 ) . |𝑆𝑟𝑒𝑐𝑖 | fused. The only case that the precision increased is when 𝐽 𝑆𝑗 ∈𝑆 recommendation from unidirectional readership SN is used with co- authorship SN recommendation. Where |𝑆𝑟𝑒𝑐𝑖 | is the number of recommenders that recommend item i. We use the cross-source hybrid if the user has relations in both social Table 3: Comparison between the user coverage of using different hybrid approaches and using recommendation from ISNs alone Hybrid with ISN data Hybrid with co-authorship only friendship SN SN Reciprocal 1.59% 2.18% 18.58% readership ISN Unidirectional 37.22% 37.43% 45.14% readership ISN Co-readership 87.25% 88.9% 92.71% ISN Figure 2: Comparison between using recommendations from ISNs only, or fusing the recommendation with co-authorship SN Tag-based ISN 85.55% 85.62% 86.57% and friendship network 5.3.4 User Coverage 6. Conclusions and Future Work While measuring the prediction accuracy of recommendation to filter We compared the prediction accuracy and the recommendation several recommendation approaches is important, it is not the only coverage in three implicit social networks: readership ISN (reciprocal way to evaluate the performance of a certain recommendation relations or unidirectional relations), co-readership ISN and tag-based approach. Non-performance measures, such as serendipity, diversity, ISN. Then we fused recommendation from these networks with two novelty, or coverage, can also evaluate recommendation approaches social networks that are based on explicit social relations between [12]. One measure that compares the capability of different users, namely: friendship SN and co-authorship networks. Weighted recommending approaches to produce recommendations for a larger sum approach was used to fuse the recommendations from two set of users is the coverage measure, which is the ratio of users who sources (implicit and explicit social networks). The experiments receive nonempty recommended sets to the total number of users. showed that fusing the recommendations from each ISN with The more coverage provided, the better the recommending algorithm. recommendations from either the friendship or co-authorship explicit We found that the co-readership ISN had the highest user coverage network is beneficial in increasing the user coverage. In addition, the (87.25 percent), higher than the tag-based ISN (85.55 percent), the prediction accuracy of all the recommendations from ISNs improved unidirectional readership ISN (37.22 percent), and the reciprocal when fused with the friendship explicit SN, but fusing with the co- readership ISN (1.59 percent). We found that only 18 percent of authorship SN did not help in improving the recommendation users have explicit social relations and the average number of social accuracy. Therefore, a hybrid approach that fuses recommendations relations per user is only 0.31. The co-authorship explicit SN has a from explicit social networks such as friendship can increase both very low coverage (1.873 percent). A tradeoff is noticed between the prediction accuracy and user coverage. This finding is beneficial for prediction accuracy and the user coverage: the more accurate the recommender algorithm designers to consider hybrid approaches that prediction, the smaller the user coverage. take into account different social relations of users. Such social relations can be harvested from bookmarking systems that allow Table 3 shows the coverage of different social networks and synchronizing information from different systems such as compares them to the coverage of the hybrid approaches. The (CiteULike.org and Mendeley.com). Our findings are also beneficial recommendation coverage increases when recommendations from for recommender systems interface designers, highlighting the need explicit and implicit SNs are combined. However, the maximum for allowing users to set weights of different recommendation coverage is reached when recommendation from the friendship SN is sources and show the recommended list with explanations. fused with any of the ISNs. This is true for all of ISNs. For example, the increase in the coverage for the reciprocal readership ISN when In the future, we want to generalize our findings by testing the the recommendation is fused with the friendship SN is more than proposed implicit social networks with other datasets and/or with 16%, while the increase in the coverage when the recommendation different applications. In this paper, we used the same weights for all fused with the co-authorship SN is only 0.59%. users to combine recommendations from different resources, which limits the personalization capabilities. In the future, we want to test Fusing recommendation from friendship SN increases both the using dynamic weights that are based on each user’s features such as prediction accuracy and the recommendation coverage. The social network features (e.g. number of incoming/outgoing social unidirectional readership ISN is the network that improved the most relations). We also want to test the recommendations produced by from fusing recommendation from friendship SN (F1-measure ISNs with real users to test the user perception of and/or satisfaction increase of almost 11%), then the reciprocal readership ISN (7.8 %), with recommendations and the degree to which users trust the tag-based ISN (2.9% increase), and finally, the co-readership ISN recommender, and also test the effect of giving the user the control (with 2.5% increase). Even though, fusing recommendation from co- on the fusing weights for explicit and implicit social network-based authorship SN did not improve the recommendation accuracy, it recommendations. improves the recommendation coverage. 7. ACKNOWLEDGMENTS This work has been supported by a PhD fellowship of the Institute of Public Administration, Riyadh, Saudi Arabia and by the NSERC Discovery Grant program. 8. References [14] McNee, S.M., Albert, I., Cosley, D., Gopalkrishnan, P., Lam, S.K., Rashid, A.M., Konstan, J.A., and Riedl, J. On the [1] Avesani, P., Massa, P., and Tiella, R. A Trust-Enhanced Recommending of Citations for Research Papers. In Proceedings of Recommender System Application: Moleskiing. In Proceedings of the 2002 ACM Conference on Computer Supported Cooperative the 2005 ACM Symposium on Applied Computing, (New York, USA, Work, (New York, NY, USA, 2002), ACM, 116–125. 2004), ACM Press, 1589–1593. [15] Mican, D., Mocean, L., & Tomai, N. (2012). Building a Social [2] Basu, C., Hirsh, H., Cohen, W.W., and Nevill-Manning, C. Recommender System by Harvesting Social Relationships and Trust Technical Paper Recommendation: A Study in Combining Multiple Scores between Users. In W. Abramowicz, J. Domingue, & K. Węcel Information Sources. J Artif Int Res 14 (2001), 231–252. (Eds.), Business Information Systems Workshops, Springer Berlin [3] Bellogín, A., Cantador, I., and Castells, P. A Study of Heidelberg. 1-12. Heterogeneity in Recommendations for a Social Music Service. In [16] Lee, D.H., and Brusilovsky, P. Improving Recommendations Proceedings of the First International Workshop on Information Using Watching Networks in a Social Tagging System. In Heterogeneity and Fusion in Recommender Systems, (New York, Proceedings of the 2011 iConference, (New York, NY, USA, 2011), NY, USA, 2010), ACM, 1–8. ACM, 33–39. [4] Bellogín, A., Cantador, I., Díez, F., Castells, & Chavarriaga, E. [17] Lee, D.H., and Brusilovsky, P. Interest Similarity of Group An empirical comparison of social, collaborative filtering, and hybrid Members: The Case Study of CiteULike. In the WebSci10: Extending recommenders. ACM Trans. Intell. Syst. Technol. 4 (2013), 14:1- the Frontiers of Society On-Line, Raleigh, NC: USA, 2010). 14:29. [18] Lee, D. 2013. Personalized Recommendations Based On Users’ [5] Bostandjiev, S., O’Donovan, J., & Höllerer, T. TasteWeights: A Information-Centered Social Networks. Doctoral Dissertation. Visual Interactive Hybrid Recommender System. In Proceedings of University of Pittsburgh. the Sixth ACM Conference on Recommender Systems, (New York, NY, USA, 2012) ACM, 35-42 [19] Liu, B. Informational Retrieval and Web Search. In Web Data Mining: Exploring Hyperlinks, Contents and Usage Data, (New [6] Burke, R. Hybrid Recommender Systems: Survey and York, NY, 2007), Springer, 183–236. Experiments. User Modeling and User-Adapted Interaction, 12,4 (2002), 331–370. [20] Liu, F., and Lee, H.J. Use of Social Network Information to Enhance Collaborative Filtering Performance. Expert Syst. Appl. 37 [7] Golbeck, J. FilmTrust: Movie Recommendations from Semantic (2010), 4772–4778. Web-Based Social Networks. In ISWC2005 Posters & Demonstrations, (2005), PID–72. [21] O’Donovan, J., and Smyth, B. Trust in Recommender Systems. In Proceedings of the Tenth International Conference on Intelligent [8] Groh, G., & Ehmig, C. Recommendations in Taste Related User Interfaces, (New York, NY, USA, 2005), ACM, 167–174. Domains: Collaborative Filtering vs. Social Filtering. In Proceedings of the 2007 International ACM Conference on Supporting Group [22] Pera, M.S., and Ng, Y.-K. A Personalized Recommendation Work, (New York, NY, USA, 2007), ACM, 127-136. System on Scholarly Publications. In Proceedings of the Twentieth ACM International Conference on Information and Knowledge [9] Guy, I., Zwerdling, N., Carmel, D., Ronen, I., Uziel, E., Yogev, Management, (New York, NY, USA, 2011), ACM, 2133–2136. S., & Ofek-Koifman, S. Personalized Recommendation of Social Software Items Based on Social Relations. In Proceedings of the [23] Pera, M.S., and Ng, Y.-K. Exploiting the Wisdom of Social Third ACM Conference on Recommender Systems, (New York, NY, Connections to Make Personalized Recommendations on Scholarly USA,2009), ACM, 53-60. Articles. J. Intell. Inf. Syst. 42 (2014), 371–391. [10] Herlocker, J. L., Konstan, J. A., Terveen, L. G., John, & Riedl, [24] Tang, J., Hu, X., and Liu, H. Social Recommendation: A Review. T. (2004). Evaluating collaborative filtering recommender systems. Soc. Netw. Anal. Min. 3 (2013), 1113–1133. ACM Trans. Inf. Syst. 22, 1 (2004), 5-53 [25] Torres, R., McNee, S.M., Abel, M., Konstan, J.A., and Riedl, J. [11] Kautz, H., Selman, B., and Shah, M. Referral Web: Combining Enhancing Digital Libraries with TechLens+. In Proceedings of the Social Networks and Collaborative Filtering. Commun ACM 40 Fourth ACM/IEEE-CS Joint Conference on Digital Libraries, (New (1997), 63–65. York, NY, USA, 2004), ACM, 228–236. [12] Ma, H., Yang, H., Lyu, M.R., and King, I. SoRec: Social [26] Yuan, Q., Zhao, S., Chen, L., Liu, Y., Ding, S., Zhang, X., & Recommendation Using Probabilistic Matrix Factorization. In Zheng, W. Augmenting collaborative recommender by fusing Proceedings of the Seventeenth ACM Conference on Information and explicit social relationships. In Workshop on Recommender Systems Knowledge Management, (New York, NY, USA, 2008), ACM, 931– and the Social Web, with Recsys 2009. 940. [13] Massa, P., and Avesani, P. Trust-Aware Recommender Systems. In Proceedings of the 2007 ACM Conference on Recommender Systems, (New York, NY, USA, 2007), ACM, 17–24.