A Research Platform for Recommendation within Social Networks Amit Sharma Dept. of Computer Science Cornell University Ithaca NY 14853 asharma@cs.cornell.edu ABSTRACT Thus, in domains where social connections overlap with user Recommendations within a network do affect, and get af- interest, social recommendation may be useful, in others, not fected by, the information flow and the social connections so much [13]. This claim is supported by evidence that rec- within the network. Thus, designing a network-centric rec- ommendations based on implicit networks constructed from ommender system requires understanding people’s prefer- domain-specific user activity give good results on predicting ences, their social connections, as well as the characteristics users’ preferences [5, 15]. of the network they inhabit. This creates a major challenge in research on network-centric recommendations—exploring However, we argue that there is more to explicit social net- questions around networks and recommendation is hard be- works that offline measures of recommender performance cause they invariably depend on the interaction between may not capture. For instance, studies with Facebook and more than one user. We describe a research platform that Google users have found that showing the name of a partic- we have built that helps us answer network-centric research ular friend with a recommendation (music or news) can alter questions. We present an overview of the system and demon- a user’s perception of it. [9,12]. Recommendations based on strate its usefulness through an example study involving di- social connections may also help users navigate their social rected suggestions between pairs of participants. As a useful network and help them become more aware of the interests side-effect, it is also helping us collect data about people’s and preference of people in it. As often happens in the offline preferences in social networks. world, friends of a user may also like to recommend items directly to him [2]. In such cases, and others such as group Categories and Subject Descriptors recommendation, these recommendations can help support shared experiences as well as influence the interpersonal re- H.3.3 [Information Storage and Retrieval]: Information lationships between people. Search and Retrieval—Information Filtering; H.1.2 [Models and Principles]: User/machine systems—Human Factors In addition, a user’s preferences are not static; they are con- tinuously being influenced by their network. The structure General Terms and properties of a network, as a whole, are also affected by Experimentation, Human Factors the connections between people and their activities within the network [4]. These factors suggest that there is value 1. INTRODUCTION in considering recommendations as embedded within a so- Recent research on recommendation using social networks cial network, rather than being served in isolation. We call has taken two main approaches—augmenting collaborative this approach to recommendation network-centric [11], in filtering with social data [7,10], or using only the first-degree contrast to the network-aware approaches described earlier. connections for recommendation [3, 6]. However, the gains reported by using social signals to recommend items are only The principles of network-centric recommendation are based slight [1], except for network-specific tasks such as friend on the observation that recommendations do affect, and recommendation. get affected by, the information flow and social connections within a network. Thus, designing a network-centric recom- One of the reasons for mixed results with network-aware mender system requires understanding people’s preferences, approaches may be that people’s explicit social connections their social connections, and the characteristics of the net- have little to do with their interests in a particular domain. work they inhabit. Framed this way, a challenge of network-centric recommen- dation is that it makes designing and evaluating systems hard, because they invariably depend on interactions be- tween more than one user. This paper describes how we can use PopCore1 , a research platform we have built, to ex- plore questions around networks and recommendation. As a network-centric system built on top of Facebook’s social 1 http://www.popcore.me Figure 1: The PopCore interface. Recommended items, occupying the center of the screen, are a mix of algorithmic recommendations and items directly suggested to a user by her friends. Sliders on the right help a user control the sources for recommendation. Visualizations on the left help in network awareness, in this case a word cloud of popular items among a user’s friends. network, PopCore serves as a research tool to support live evaluations of network-centric recommendations, and collec- tion of user reactions and feedback in a real network setting. It is also a functional recommender system in its own right, much like the MovieLens system2 . 2. OVERVIEW OF SYSTEM DESIGN Figure 1 shows a screenshot of PopCore, which we first pro- posed two years ago in this workshop [14]. It uses Facebook as the underlying network and covers the entertainment do- Figure 2: Possible actions on a recommended item, main, including movies, music, books and television shows. shown when a user clicks on it. A user may Like it on Facebook, rate it, recommend it to a friend or The interface has three main parts: recommendations in add it to her personal queue. the center, visualizations on the left and user controls on the right of the screen. Recommended items are a mix of recom- mendations computed from items Liked by a user’s friends’, list of recommended recipients who PopCore predicts might and directed suggestions from his/her friends. These di- be interested in the item (based on the similarity between rected suggestions are one of the key features of the system— the item and the user’s friends). items can be recommended manually by users to their friends. These suggestions serve two purposes. First, they utilize Visualizations, such as the item cloud of most popular items friends’ knowledge to bring interesting recommendations. Liked by a user’s friends and updates of recent activity on Second, they allow people to express their preferences to the app (Figure 1), help create increased network awareness their friends and support conversations and shared experi- for the user. User controls on the right, which allow a user to ences on the recommended items. tailor recommended items based on their popularity, social closeness or similarity of the people connected with those For each recommendation, a user may Like it on Facebook, items, also help users find interesting content and navigate rate it, recommend it to some friends or add it to a personal their friends’ interests. queue (Figure 2). When a user clicks on the Recommend button for an item, he can choose one or more of his friends to recommend the item to. For convenience, we also offer a 3. POPCORE AS A RESEARCH PLATFORM We now describe how we are using the system as a test bed 2 http://www.movielens.org for research. In addition to the main website, we have a created a separate labs website3 . It runs on the same core standing of the receiver’s preferences. Understanding the infrastructure, but has custom code that allows us to run processes behind directed suggestions, specifically around experiments. Participants are asked to sign up and give item and receiver selection, can help design recommenda- permission for each experiment separately. tion systems that support such suggestions. The core infrastructure of PopCore supports an ego-centric One of our first goals is to simply compare the recommenda- view of the network for each user, collecting Likes and net- tion quality of directed suggestions versus algorithmic rec- work information about the user and his friends that can ommendations. In the following paragraphs, we present how be used for a variety of studies. In our first experiment, the PopCore platform can be useful for investigating this we showed users a variety of different recommendation algo- question. rithms, some of which used past Likes of the users’ friends. Users were asked to rate and Like items and react to the idea Experiment Design: Previous research on helping people of network-centric recommendation [11]. We found that an share news items in RSS feeds suggests that manual recom- algorithm suggesting the most popular items among a user’s mendations to friends can be useful [2]. Our approach is friends performed the best among those that used ego net- somewhat like Krishnan et al., who asked people who did work information. It was also significantly better rated than not know a target user to make recommendations based on an algorithm based on overall network popularity. Users’ re- the target user’s ratings [8]; however, instead of providing a actions to the recommendations confirmed that network ties list of ratings from a stranger, we ask people to make rec- can provide a useful way to choose potential content and ommendations for their friends based on what they already neighbors for information filtering tasks. know about them. In the study described above, participants did not know A simple way to compare the quality of recommendations the social nature of the recommendations. Subsequently, we would be to analyze logs of user ratings in PopCore and have used the platform to study how presenting a social ex- compare the performance of directed suggestions and algo- planation along with a recommendation can affect user rat- rithmic recommendations. However, this comparison will ings for music [12]. Here, we showed users different kinds of depend greatly on the particular recommender algorithm be- explanations involving their friends, such as “X and 3 other ing used. One way to control for the recommendation algo- friends like it”, along with a recommended musician. Our rithm is to allow only those directed suggestions that are also results show that social explanation has only a secondary recommended by the algorithm we are comparing against. effect on the ratings; the primary effect is that of a user’s Then, among the algorithmically recommended items, we expectation of liking an item. Based on the findings, we would be able to compare the ratings between items that provided a generative mixture model for a user’s decision were manually suggested and items that were not. Such a process on a recommendation. comparison would point to the quality of manual and algo- rithmic recommendations, as perceived by the user. Besides live experiments, data collected for the above stud- ies helps us reason about the properties of networks and For ease of logistics, we design a movie recommendation ex- how they might affect recommendation. In one such study, periment involving pairs of participants. The system gen- we compared offline performance of algorithms using prefer- erates its top 10 recommendations for each participant in ences of friends or the whole network, and found that algo- a pair. It then randomly shuffles the 20 recommendations rithms using just the friends’ information give comparable for a pair into a combined list, and shows the combined list results to those using the full network [13]. Further, the to each participant. Similar to the main PopCore interface, data helped us investigate how the prevalence of locality of participants are free to rate them or suggest them to their preference (concentration of Likes for an item in parts of the partner. We show this combined list to participants so that network) roughly correlates with the performance of friends- both of them rate and suggest from the same set of recom- only algorithms on three different domains, suggesting that mendations. This is done to ensure that both partners can this locality is an important phenomenon and resource for choose to start the study at different times and still rate each recommender systems. other’s potential suggestions in a single session. In addition, comparing ratings for the same item from both sender and Till now, we have described how PopCore can be a useful receiver could tell us whether people tend to recommend platform for experiments involving a user and social data. items that they themselves rate highly, or those that may However, the system is designed to support participation be relevant based on the other’s preferences. and interaction between multiple users. We now describe an ongoing experiment as an example of how PopCore can Experiment Flow: We run the study using the labs ver- be a useful tool for conducting experiments with more than sion of PopCore. In the first stage, a participant invites one participant. one of his friends as his partner. Once the other partici- pant accepts the invitation, she enters the experiment and rates/suggests 20 items. The first participant then gets an 3.1 Directed suggestions: A paired experiment email and he can go ahead and rate/suggest the same 20 We consider the practise of sharing items between people, items. During the experiment, we highlight the other par- more commonly known as “word of mouth”. Intuitively, sug- ticipant’s name whenever a user wishes to make a directed gesting an item to another person may involve two factors: suggestion (Figure 3), in order to make the directed sugges- having an opinion about the item, and having an under- tions easy to send out. Thus, at the end of the experiment, 3 we receive ratings for the items from both participants and http://labs.popcore.me we also hope to share the data we are collecting, keeping in mind the potential challenges around user privacy in gener- ating those datasets. 5. ACKNOWLEDGMENTS This work was supported by the National Science Founda- tion under grants IIS 0910664 and IIS 0845351. 6. REFERENCES [1] O. Arazy, N. Kumar, and B. Shapira. A theory-driven design framework for social recommender systems. In Journal of the Assoc. for Info. Sys., 2010. [2] M. S. Bernstein, A. Marcus, D. R. Karger, and R. C. Miller. Enhancing directed content sharing on the Figure 3: The interface for directly suggesting web. In Proc. CHI, pages 971–980, 2010. movies to a friend. In the live system, we would [3] J. Chen, R. Nairn, L. Nelson, M. Bernstein, and show a list of friends who the system predicts might E. Chi. Short and tweet: Experiments on be good to share with; for the experiment, we alter recommending content from information streams. In this behavior to show only the name of their exper- Proc. CHI, 2010. imenal partner. [4] D. Crandall, D. Cosley, D. Huttenlocher, J. Kleinberg, and S. Suri. Feedback effects between similarity and social influence in online communities. In Proc. KDD, a list of the items that they shared to one another. Par- 2008. ticipants are also presented with a questionnaire that asks [5] J. Golbeck and J. Hendler. Filmtrust: movie about how close they are to the other participant, and the recommendations using trust in web-based social reasons they receive or send recommendations in general. networks. In Proc. IEEE Consumer Comm. and Networking Conf., 2006. Note that participants have no way of detecting which items [6] I. Guy, M. Jacovi, A. Perer, I. Ronen, and E. Uziel. were suggested to them by their partner. Thus, the ratings Same places, same things, same people? Mining user can be assumed to pertain to the quality of the recommen- similarity on social media. In Proc. CSCW, CSCW dation, rather than the social connection between the sender ’10, pages 41–50, New York, NY, USA, 2010. ACM. and the recipient. [7] I. Konstas, V. Stathopoulos, and J. M. Jose. On social networks and collaborative recommendation. In Proc. Initial Results: We are currently in the process of col- SIGIR, 2009. lecting data. We are recruiting participants from a mix of [8] V. Krishnan, P. K. Narayanashetty, M. Nathan, R. T. college student pools and Amazon mechanical turk. From Davies, and J. A. Konstan. Who predicts better?: an analysis of the initial data on 27 pairs, we find that items results from an online study comparing humans and which were manually suggested have a higher average rating an online recommender system. In Proceedings of the than those that were algorithmically recommended. This 2008 ACM conference on Recommender systems, 2008. supports the notion that manual recommendations are use- [9] C. Kulkarni and E. Chi. All the news that’s fit to ful, and in some cases, better than an algorithmic recom- read: a study of social annotations for news reading. mendation. However, it may also be the case that people In Proc. CHI, 2013. tend to recommend highly popular items which are more [10] H. Ma, D. Zhou, C. Liu, M. R. Lyu, and I. King. likely to get higher ratings. We hope to use offline data to Recommender systems with social regularization. In corroborate these findings. Proc. WSDM, 2011. In addition, we find that the average rating for manual rec- [11] A. Sharma and D. Cosley. Network-centric ommendations is higher for the senders than receivers. This recommendation: Personalization with and in social result suggests that users may be sharing items that they networks. In Proc. IEEE SocialCom, 2011. like more frequently than those which they think can be [12] A. Sharma and D. Cosley. Do social explanations relevant for the recipient. work? Studying and modeling the effects of social explanations in recommender systems. In Proc. WWW, 2013. 4. CONCLUSION [13] A. Sharma, M. Gemici, and D. Cosley. Friends, We have presented PopCore, a research platform for con- strangers, and the value of ego networks for ducting user experiments with recommendations in social recommendation. In Proc. ICWSM, 2013. networks. Our goal for this paper is to show how the system [14] A. Sharma, M. Malu, and D. Cosley. Popcore: A has helped us conduct experiments that require social data system for network-centric recommendations. In Proc. and/or active participation of users embedded in a network. RSWEB Workshop, RecSys, 2011. We envision it as an open platform for research, for which [15] T. Zhou, J. Ren, M. Medo, and Y.-C. Zhang. we would like to discuss potential ways of collaboration at Bipartite network projection and personal the workshop. For instance, one of the ways is to allow other recommendation. Phys. Rev. E, 76, Oct 2007. researchers to setup experiments on PopCore. In addition,