NETWORK SCIENCE vs WIKILEAKS etc. Lovro Subelj Cookie seminar at - PowerPoint PPT Presentation

NETWORK SCIENCE vs WIKILEAKS etc. Lovro ˇ Subelj Cookie seminar at FRI 19.3.2015

ACKNOWLEDGEMENTS Aleksandar Angelina Peter Marko Slavko Marinka Julian

KARATE CLUB NETWORK 17 16 23 15 6 31 19 7 11 9 21 33 5 14 34 10 30 2 20 18 1 27 3 24 22 32 29 4 8 28 12 13 26 25 Zachary, W. W., J. Anthropol. Res. 33 , 452–473 (1977).

KARATE CLUB SPLIT 17 16 23 15 6 31 19 7 11 9 21 33 5 14 34 10 30 2 20 18 1 27 3 24 22 32 29 4 8 28 12 13 26 25 Zachary, W. W., J. Anthropol. Res. 33 , 452–473 (1977).

NETWORKS WITH COMMUNITIES Cross Five Tennessee 17 Vau Louisiana State Kentucky Georgia Florida 16 Memphis Southern Mississippi 23 Mississippi State Trigger MN60 Georgia Tech Vanderbilt Louisville Wake Forest Clemson 15 South Carolina Arkansas Alabama Birmingham SMN5 Mississippi MN83 CCL Zap Notch 6 East Carolina 31 Duke Tulane Haecksel Maryland Louisiana Monroe 7 19 MN105 Topless Mus Florida State Houston Virginia Jonah Quasi North Carolina State Auburn Alabama Louisiana Lafayette Patchback Middle Tennessee State Army Connecticut 11 Number1 Central Florida North Carolina Cincinnati 9 21 Double MN23 Oklahoma State 33 SN100 SN89 Arkansas State Kansas TR99 Virginia Tech Fork TSN103 SN9 Syracuse New Mexico State Missouri Grin Navy Beescratch Boston College Oklahoma 5 Scabs Jet Temple Louisiana Tech Baylor Texas 14 10 34 SN4 Southern Methodist 30 TR82 Eastern Michigan West Virginia Kansas State North Texas Stripes Ball State Texas Tech 2 Kringel Beak SN90 DN63 Rutgers 20 Hook Oscar Northern Illinois Brigham Young Iowa State Tulsa Indiana Nebraska Texas A&M Web Buffalo Akron Pittsburgh Miami Florida Boise State Miami Ohio 18 Texas Christian Texas ElPaso 1 TR120 Shmuddel Fish Upbang Rice SN63 Nevada Las Vegas 27 Northwestern Minnesota Colorado 3 Feather Notre Dame Air Force Utah State TR77 Gallatin Marshall Wyoming Hawaii SN96 Knit Toledo Central Michigan 24 Bowling Green State Ohio Idaho 22 TSN83 DN21 Iowa Michigan State New Mexico San Jose State 29 Zipfel Thumper Nevada 32 PL Western Michigan Wisconsin Kent Penn State Fresno State TR88 DN16 Illinois Colorado State 8 4 Purdue Michigan 28 Utah Bumper UCLA Whitetip Ripplefluke 12 Ohio State San Diego State Stanford Wave Southern California Washington Washington State 13 Oregon 26 Arizona State Zig California 25 Oregon State Arizona Karate club Bottlenose dolphins American football Synthetic graph Random graph

NETWORK COMMUNITY DETECTION • hierarchical clustering, • modularity optimization, • statistical inference, • spectral methods, • map equation, • dynamics etc. Girvan, M. & Newman, M. E. J., P. Natl. Acad. Sci. USA 99 , 7821–7826 (2002).

LABEL PROPAGATION ALGORITHM 17 16 23 15 6 31 19 7 11 9 21 33 5 14 34 10 30 2 20 18 1 27 3 24 22 32 29 4 8 28 12 13 26 25 Raghavan, U. N., Albert, R. & Kumara, S., Phys. Rev. E 76 , 036106 (2007). ˇ Subelj, L. & Bajec, M., Phys. Rev. E 83 , 036103 (2011) etc.

COMMUNITY DETECTION IN PRACTICE Hric, D., Darst, R. K. & Fortunato, S., Phys. Rev. E 90 , 062805 (2014). Hric, D., Proceedings of NetSci ’14 , Berkley, CA, USA (2014).

COMMUNITY DETECTION TASKS EXPLORATORY TASK PREDICTIVE TASK TRAINING SET TEST SET TRAINING SET TEST SET ˇ Subelj, L., Proceedings of ARS ’15 , Capri, Italy (2015). ˇ Subelj, L., submitted at NetSci ’15 , Zaragoza, Spain (2015).

APS & WIKILEAKS NETWORKS APS WikiLeaks DATA 1893-2013 1966-2010 citation reference 526,527 papers 52,416 cables NETWORK 5,989,263 citations 78,506 references 12 journals 3 privacy levels CLUSTERS 301 sections 263 embassies 14 algorithms 26 algorithms SETTING 1893-2012 1966-2009 TRAINING 2013 (4%) 2010 (17%) TEST

APS & WIKILEAKS RESULTS NORMALIZED MUTUAL INFORMATION EXPLORATORY TASK 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 12 journals 301 sections 3 privacy levels 263 embassies PREDICTIVE TASK CLASSIFICATION ACCURACY 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 12 journals 301 sections 3 privacy levels 263 embassies SPEARMAN CORRELATION TASK CORRELATION 1 0.75 0.5 0.25 0 −0.25 −0.5 −0.75 −1 12 journals 301 sections 3 privacy levels 263 embassies

YOUTUBE, DBLP & JAVA NETWORKS YouTube DBLP java DATA social collaboration software 39,841 users 317,080 authors 2,378 classes NETWORK 224,235 friends. 1,049,866 collabs. 14,619 dependens. 12,986 groups 98,326 venues 54 packages CLUSTERS 14 algorithms 14 algorithms 26 algorithms SETTING leave-one-out leave-one-out leave-one-out

YOUTUBE, DBLP & JAVA RESULTS NORMALIZED MUTUAL INFORMATION EXPLORATORY TASK 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 12,986 categories 98,326 venues 54 packages PREDICTIVE TASK CLASSIFICATION ACCURACY 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 12,986 categories 98,326 venues 54 packages SPEARMAN CORRELATION TASK CORRELATION 1 0.75 0.5 0.25 0 −0.25 −0.5 −0.75 −1 12,986 categories 98,326 venues 54 packages

FRI COLLABORATION NETWORK Web of Science (-2014).

NETWORKS ARE FUN (for BSc & MSc theses write to lovro.subelj@fri.uni-lj.si )