Snag’em: Graph Data Mining for a Social Networking Game Veronica Cateté Drew Hicks North Carolina State North Carolina State University University 911 Oval Drive 911 Oval Drive Raleigh, NC 27606 Raleigh, NC 27606 vmcatete@ncsu.edu aghicks3@ncsu.edu Collin Lynch Tiffany Barnes North Carolina State North Carolina State University University 911 Oval Drive 911 Oval Drive Raleigh, NC 27606 Raleigh, NC 27606 cflynch@ncsu.edu tmbarnes@ncsu.edu ABSTRACT foundational networks and thus face difficulties making con- New conference attendees often lack existing social networks nections. Based on Tinto’s Theory of University Departure, and thus face difficulties in identifying relevant collaborators increased interaction with other students, faculty, staff and or in making appropriate connections. As a consequence community supporters can increase the retention rate of mi- they often feel disconnected from the research community nority populations and sense of community within secondary and do not derive the desired benefits from the conferences and post-secondary academic communities [7]. that they attend. In this paper we discuss Snag’em, a social network game designed to support new conference attendees In academia, sense of community has a strong positive cor- in forming social connections and in developing an appropri- relation with retention [7]. Research indicates that students ate research network. Snag’em has been used at seven pro- who do not feel as if they are part of a larger academic com- fessional conferences and in four student settings and is the munity are less likely to participate in extracurricular activ- subject of active research and development. The developers ities and organizations. This leads to lower retention rates, have sought to make the system engaging and competitive especially amongst minority students who suffer without a while preventing players from ‘gaming’ it and thus accru- strong student support group [7]. A feeling of community ing points while neglecting to form real-world connections. can be nurtured with small group activities that augment We briefly describe the system itself, discuss its impact on the individual’s role within a setting and helps students to users, and describe our ongoing work on the identification foster connections [8]. of critical hub players and important social networks. Snag’em was designed as a pervasive game to encourage valuable professional networking and promote sense of com- Keywords munity. The system’s pervasive features are designed to Social Networks, Gamification, Conferences, Underrepresented help players translate their in-game networks directly into Populations real world peer groups. The system was originally created for the 2009 Students and Technology Academia Research & 1. INTRODUCTION Service (STARS) conference. This conference is unusual in Social networking is an essential task at any academic con- that it is an academic conference designed specifically to en- ference or professional venue. One of the primary goals of gage with minority and female undergraduates majoring in attendees is to seek out relevant work, identify potential computing fields. Students who attend the conference par- collaborators, and to maintain existing connections. Many ticipate in competitions and attend training sessions to sup- of these contacts are made by building upon existing re- port engagement and research. Studies conducted at prior lationships and by expanding the attendees existing social conferences has shown that while students were engaged in network. New conference goers however, particularly stu- the training sessions and vigorously involved in learning they dents and historically underrepresented groups, lack these did not develop the lasting social connections that can arise out of conferences. Snag’em was designed to engage stu- dents in social networking through gamification of the pro- cess. Prior research has shown that social games can help people to engage in otherwise challenging or uncomfortable situations [6, 4, 2, 3]. Snag’em functions as a large human scavenger hunt. Play- ers are assigned a set of relevant tags (e.g. “I’m a games researcher”, or “I’m interested in data-mining”). They are Figure 2: Here is an example of a message sent in Figure 1: The browser interface for mission assign- game after a conversation between players. ments. Snag Snapshots highlight missions recently completed. To date, Snag’em has been used at seven academic confer- ences. It has also been deployed to help incoming freshman then assigned a set of missions (e.g. “Find someone who spe- and transfer students connect at four academic institutions. cializes in HCI”) which they must complete by identifying In 2009, for example, Snag’em was used by new students in and engaging with an appropriate individual. The system the College of Computing and Informatics at the University was developed in PHP with a MySQL backed and provides of North Carolina at Charlotte. Students were able to play a web-based front end for players to edit their profile and to the game during the freshman orientation week with kiosks record interactions. We have also developed a mobile ver- available for students to sign up located in the College of sion of Snag’em which allows players to access the game via Computing and Informatics. SNAG’EM was used alongside tablets and smartphones. The game itself is designed for other social activities to get students acquainted with each easy deployment to new conferences and we are presently other, the faculty, and the CCI campus. adding features that will allow us to automatically populate the database with initial tags. 2. PRIOR ANALYSIS Figure 1 shows a snapshot of the mission browser screen from We have studied the impact of Snag’em on users and found the web version of Snag’em. Contact is registered when the that playing the game improved conference attendees’ sense players enter a 4-digit ID from the other person. In addi- of community [6, 1]. We have also analyzed the existing tion to missions the systems also allows players to record dataset both to test the implementation of the Snag’em fea- notes about one-another for future reference (e.g. “I should tures, and to identify hubs or critical players whose activity e-mail my proposal to him after the conference”) and to send predicts the behavior of others. one-another messages. A sample message from the mobile interface is shown in Figure 2. Snag’em can also be con- In analyzing the game mechanisms we have focused primar- figured to suggest specific individuals that students should ily on the STARS 2009 dataset. As mentioned above STARS make contact with based upon their mutual interests or so- is primarily targeted at undergraduate students specifically cial connections. females and underrepresented minorities. We deployed the system via the conference infrastructure and set up a table The system logs all player interactions including tag up- near the registration booth. The game was active during dates, missions completed, notes made, messages, sent, con- the first two full days of the conference. The conference nections added, and so on. This provides a rich dataset of had 280 attendees 60.0% of whom were female (N=168) and information that we can use to analyze social patterns at 70% of which (N=196) were students. Roughly 28% of the conferences and to improve the impact of the intervention. conference-goers played the game (N=80) of whom 50% were In addition to the raw logs the game contains a number of female. In previous analysis 35.0% of the players were clas- features to support easy analysis. The developers have cre- sified as active. It is important to note that this data was ated a set of badges that allowed administrators to easily collected on an earlier version of SNAG’EM where players track the number of people playing via the mobile or web could snag each other only once, and only a single mission interfaces as well as the number of missions completed. The was available at a time. Because completing missions was badge system also provides a simple visual record of the significantly more difficult in this version of the game, play- types of features (i.e. notes, tags, avatars) each player is us- ers were classified as active if they completed at least two ing. The badge systems also allows administrators to note missions. An additional 50% of the players were classified the frequency of use, time of day that players are online and as Interested, meaning they did more than just register for so on. the game or that they completed one mission. Figure 4: Correlation between active player hubs Figure 3: Visualization of community center 4142, and number of interactions. with one of that user’s maximal cliques highlighted. likely to engage in the deep and meaningful conversations Our analysis of this data was focused primarily on the mis- required or to form lasting connections. sion and scoring systems. In 2009 the mission system was relatively simple and focused solely on guiding students to In response to these results we have overhauled the scor- locate a single individual with a desired tag. Players were ing system. This included changing the connectivity bonus then guided to record the match via the ID system discussed to reward players based upon the size of the largest clique above. Both the missions generated and points received were that they participate in. Players are now rewarded more determined by the state of the current network. When gen- for expanding this clique, thus deepening their social net- erating missions we attempted to ensure that they were of works, than they are for adding an unrelated individual to varying difficulty, and were relevant to the current user. In their friends of friends. We have also allowed players to re- this iteration of the system the missions could only be sat- snag the same individual for multiple missions with a low isfied by identifying someone whom the user had not previ- penalty for re-snags, and have begun to reward players with ously snagged. The target tags were selected from the full set points for allowing themselves to be snagged to help others listed in the system. Easy missions were assigned high fre- complete a mission. We have not yet analyzed the effects of quency tags (more than 12 of the non-adjacent users), while these changes on a the dataset. medium missions were assigned tags that are present in 14 of non-adjacent users and hard missions required tags present We have used two measures of importance when identifying in less than 14 of the non-adjacent community. critical players. The first is the simple interaction frequency as measured by the number of outgoing arcs from a player in The difficulty of the mission determined the base score which the network. The second is membership in maximal cliques, was then modified by a connectedness factor. This factor that is, cliques which are not part of a larger clique. Play- was greater than 1 if adding this connection expanded your ers that participate in a large number of maximal cliques “Friends of friends,” that is, the number of vertices less than are hubs. We were able to identify three distinct user com- 2 edges distant from the user. The connectedness factor was munities in the STARS 2009 dataset that centered on these less than 1 if you completed the mission using the ID of a hubs. A sample community graph is shown in Figure 3. We person you were already adjacent to, In this way we hoped also found that the activity of these hub players was highly to encourage players to branch out. correlated with the activity of the other players in the com- munity (r=0.827). A graph of these spikes is shown in Figure When developing the system we had hoped that players 4. More specifically, on any day where one or more of the would develop social networks that exhibited breadth (i.e. hub players were active, we observed spikes in the number meeting lots of people), depth (i.e. getting to know some in- of interactions taking place across users. We were able to dividuals well), and mutuality (i.e. snags in both directions). observe a similar effect (r = 0.659) on days when the devel- We therefore hoped that users’ immediate neighborhoods opers had a booth/kiosk available. would be large and relatively dense with multiple snags be- tween some people and bidirectional connections. When an- We also performed an analysis of hub players using the alyzing the STARS 2009 dataset, however, we found that UNCC Student Orientation dataset described above. In this this was not the case. Rather the game mechanics encour- dataset 91 of the 1290 potential students registered to play aged players to make a relatively large number of unrelated Snag’em of which 22% (N=20) were female [5]. This data connections which, in turn, produced relatively broad and was collected on a version of Snag’em permitting multiple shallow social neighborhoods with very few inbound arcs. In missions and allowing players to connect with the same user fact some players actually opted to hide their IDs so that no multiple times.We classified players as active if they com- other player could gain points by using them to complete a pleted 5 or more missions. In total, 9 users were active mission. As a consequence the attendees were actually less users during this study. However, all of these players were moderators or members of the development team. In this 4. ACKNOWLEDGMENTS deployment almost all of the game interaction took place This research was supported by the NSF GRFP Fellowships at the registration table thus making the administrators re- No. 0900860 & No. 1252376 and BPC Grant No. 0739216 sponsible for most of the activity. We had hypothesized and No. 1042468 Thanks to all developers who have worked that the moderators would only need to initiate the game on the SNAG’EM project. The authors also wish to thank and then it would be self-sustaining. As our analysis shows Shaghayegh Sahebi for her expert advice. however, this was not the case. In general the players did not think about the game outside of the advertised area. 5. REFERENCES [1] S. L. Finkelstein, E. Powell, A. Hicks, K. Doran, S. R. 3. OPEN QUESTIONS & FUTURE WORK Charugulla, and T. Barnes. Snag: using social Our prior research has focused on identifying key players networking games to increase student retention in using graph methods. We plan to continue examining these computer science. In Proceedings of the fifteenth annual key players in future work and to modify the mission se- conference on Innovation and technology in computer lection criteria to better engage players that have not been science education, pages 142–146. ACM, 2010. active recently. Our chosen method of community detec- [2] M. Montola. Exploring the edge of the magic circle: tion, based upon maximal cliques, is both computationally Defining pervasive games. In Proceedings of DAC, page expensive on large networks and can change substantially 103, 2005. based upon small shifts in the network. Using a simpler, [3] M. Montola. A ludological view on the pervasive less volatile measure to identify community centers would mixed-reality game research paradigm. Personal and allow us to adapt the gameplay based upon those communi- Ubiquitous Computing, 15(1):3–12, 2011. ties more efficiently. This would in turn enable us to encour- [4] E. Powell and T. Adviser-Barnes. A framework for the age new players to specifically seek out these active players design and analysis of socially pervasive games. 2012. in an effort to better engage them from the start. Differ- [5] E. Powell, F. Stukes, T. Barnes, and H. R. Lipford. ent community detection algorithms might identify different Snag’em: Creating community connections through hub players, or provide different ways of scoring missions games. In Privacy, security, risk and trust (passat), that help to foster larger communities. Further develop- 2011 ieee third international conference on and 2011 ment in this area might facilitate play in the absence of an ieee third international conference on social computing instigating ‘active player’ or outside of areas with an active (socialcom), pages 591–594. IEEE, 2011. game station or kiosk. [6] E. M. Powell, S. Finkelstein, A. Hicks, T. Phifer, S. Charugulla, C. Thornton, T. Barnes, and One open question is how to better identify hub players dur- T. Dahlberg. Snag: social networking games to ing the game, and modify mission selection criteria to engage facilitate interaction. In CHI’10 Extended Abstracts on inactive players or players who don’t need motivation to net- Human Factors in Computing Systems, pages work. These ‘social elites’ are important to attract, as they 4249–4254. ACM, 2010. are precisely who we should be encouraging our players to network with. If we are better able to build and analyze our [7] V. Tinto. Taking Student Retention Seriously: networks, we may be able to offer features to these social Rethinking the First Year of College. NACADA elites that would attract them to Snag’em as a system more Journal, 19(2):5–10, 2000. than the gamification aspects would. We hope to explore [8] S. White. Algorithms for estimating relative techniques for reliably generating edges and tags for users importance in networks. Proceedings of the ninth ACM based on existing data sources like conference proceedings SIGKDD international, pages 266–275, 2003. or citations. This would reduce the burden of entry on new players, particularly elites, and make it more likely for those users to participate in networking (if not gameplay) using SNAG’EM. We also plan to expand our in-game evaluation of Snag’em itself. We are presently adapting the system to poll play- ers for their opinions as the system is used. This will bet- ter help us to identify the immediate impact of the system on users’ social connections. We will be deploying some of these new features of the system during the 2014 Educa- tional Datamining Conference in London as well as subse- quent conferences in 2014 and 2014.