Michael, Great if not an awesome question!! Let me try my best to assist you in finding more information regarding your questions. Q1 - "After the fact" A1 - First, the government is CURRENTLY using open source and classified data at their disposal to map social networks of terrorists and potential terrorists. Unfortunately, most of this information is classified, so the details of an active operation are not available to the average citizen. Therefore, most of the information is declassified after the information is already available to the public (or leaked via the press). We have learned of the 9/11 cells after 9/11 and through many academics picking apart the network. A great book on this subject is by Marc Sageman titled, "Understanding Terror Networks" (2004), ISBN-13: 9780812238082. Two examples of social network analysis programs that were attempting to unearth information on terror networks is the project called Able Danger [1] and, more recently, the revelations surrounding the alleged Ft. Hood gunman, Major Nidal Malik Hasan and his ties to radical cleric Anwar al Awlaki [2]. Although we use the term Social Network Analysis in defining the connection of people with one another, sometimes the Intelligence Community may refer to such operations as Information Operations or IO) or Data Mining, etc. However, the government is currently using such approaches both in the Afghanistan and Iraq theater of operations to unearth various insurgent or terrorist cells. You can actually go to USAJobs.gov or Indeed.com and find current openings for government and government contractors for Social Network Analysis positions that are either CONUS or in theater. I did a search today and found several positions specifically posted by a variety of government contractors, including: SAIC, Booz Allen Hamilton, and BAE Systems (all DOD government contractors). Sources: [1] http://en.wikipedia.org/wiki/Able_Danger [2] http://abcnews.go.com/Blotter/official-nidal-hasan-unexplained-connections/story?id=9048590 Q2 - "Unobservable Variables" A2 - Unobservable variables are sometimes limited by the scope of data you have for a data set and therefore limit what you can map (observe). As with Able Danger (and as I understand the program), the government had classified information on a variety of known terrorists and tried to use open source data to map out a hidden network based on merging the two data types. If you have variables in both data sets that can be linked together via a key, you may be able to unearth hidden relationships. On another note, the government is not working with a data set that we, as civilians, without classified clearances, can see. Therefore our data set is limited, whereas there's is more voluminous and therefore they have the ability to unearth many more edges than we can. Keep in mind that with the Ft. Hood suspect, he was found to be in contact with a radical cleric in Yemen and therefore became a blip on the government's radar. Knowing that there is an edge between Hassan's node and Awlaki's node provides a revelation to someone that is also noticing his radical posts online (again, another potential revealed edge through open source data mining). His social network therefore begins to reveal relationships. Apparently, based on what I've read and interpreted, the government was either incapable of seeing the relationship (in time) or decided to overlook it (for what appears to be a variety of issues, if not due to the fear of singling out an Islamic officer in the military - time will tell what the real reason is). Obviously, if a person stays off of the "grid" and limits his data and personal interaction as much as possible, someone may not be seen in a terror network, even with multiple tools in hand - that's why it is so important that human observations (HUMINT), electronic observations (SIGINT, ELINT, etc.) and MASINT (look it up because it's challenging to explain through a post) be sifted through by a social network analyst/data specialist to try to uncover relationships and hopefully reveal terror networks before it's too late. A3 - Technically, I could actually create a data application, crawl a site (or many sites) associated with ties to terrorist-minded individuals who speak/type in a known language ((English, Farsi, Urdu, Pashtu, Arabic, etc.) and start mapping social structure with nodes related to IPs, user ids, etc. and map out a social network, simply by using open source data. If I had a separate (let's say classified) data set, I could begin to match people to IPs, user ids, pictures, nodes, locations, etc. and then begin to reveal structure and potentially determine who leads the network, what the pecking order is, what standard operating procedures are and perhaps even infiltrate and dismantle it from the inside, or use it to expand your knowledge of a larger social network connected to it. The iterations are nearly endless if you have a large enough data set! Further Sources to Consider: Valdis Krebs, Social Network Analysis Cases [Examples], http://orgnet.com/cases.html Valdis Krebs, Social Network Analysis of the 9-11 Terrorist Network, http://www.orgnet.com/hijackers.html Valdis Krebs, Uncloaking Terrorist Networks, http://firstmonday.org/htbin/cgiwrap/bin/ojs/index.php/fm/article/view/941/863 Steve Ressler, Social Network Analysis as an Approach to Combat Terrorism: Past, Present, and Future Research, http://www.hsaj.org/pages/volume2/issue2/pdfs/2.2.8.pdf. Naohiro Matsumura, David E. Goldberg, Xavier Llora, Mining Social Networks in Message Boards, http://www.slideshare.net/xllora/mining-social-networks-in-message-boards

"Can graph theory provide insight into probability of structure?" Network statistical models attempt to do just that - check out exponential random graph models or (for longitudinal network data) Tom Snijders actor-oriented models. It's still early days but we are making headway. "Network structure sensitive to error" - this can be an issue, for sure, but what it does mean is that you need statistical approaches that give you reasonable guides to uncertainty - eg through standard errors on parameter estimates, or Bayesian approaches. It may then be possible to simulate models based on a plausible distribution of estimates to check out the implied range of possible network outcomes. In terms of the value of "after the fact" data in covert networks, have a look at some of Carlo Morselli's analyses of criminal networks to see how data after the event may still provide insights into how such networks may operate. Garry Robins

I highly recommend looking at the research conducted by Kathleen Carley; she does a lot of research on terrosim networks and may have some insights to offer. You can also look up the longitudinal analysis work done by George Barnett for time series and predictability in networks.

I second that question, on an even broader level. I'm a behavioral ecologist and I'd like to make similar comparisons for groups of organisms. A particular problem I have is that network structure is so sensitive to error. So given that, and given the question posted above, I would really like to know whether graph theory + network statistics can be used predictively. Or even if it's possible to compare two different network structures to each other (or a special case of this: how the same network changes through time). Second, I have seen some studies in the animal literature that manipulate networks in computer simulations to assess the effect on the overall network structure (e.g. by knocking out individuals with certain high attribute values). Typically this results in dramatic changes, but I'm not sure such outcomes are valid since in reality one would expect other individuals to take their place and network reorganization as a result. This is quite different to an approach in which one actually experimentally removes subjects and then measures what happens to network dynamics. I am very concerned about this. I understand that this is not a typical use of the apparatus, but to me this such pursuits seem potentially useful. Thanks.