Call Graph Challenge (Missing Edge)

In my data mining class last fall we had a challenge problem that was to find the missing edge of a call graph.  We were provided with the anomalous telecommunication information of two cities, Moria and Standelf (these two cities in middle earth corresponding to two cities in the US - we were told at the end but I didn't write it down). Namely (for an entire month) we were provided with the days (number of distant days that two nodes communicated during the month), calls (number of distinct calls that the nodes made during the month), secs (the commutative length of all calls), and texts (the number of distinct texts) and our job was to predict the probability that two nodes within the graph would comunicate with each other. With the news about Verizon (or other telecoms - I haven't really paid attention) providing the NSA with similar information I thought this project was really cool.

Example call graph

Our group ended up using my  neural net, which was actually a little bit worse than to tossing a coin to decide if a link existed. I know other groups did much better (maybe getting up to 70-80%) but I didn't really pay attention. My real problem was compute power, I was using NetworkX (LANL network code, in python on a single node) and it was pretty slow to run.

I just wonder how useful this information is to the NSA. Say that they can classify 90% of the missing edges (a call between party A and B correctly). With millions of calls, 90% still leaves hundreds of thousands to sort out by other techniques.* Look at the calls distributions  lots of people make lots of calls. Look at the call duration and texts; most of the people make short calls and send out lots of texts. ** Most of the people don't leave much of a footprint to go off of.

I think the long tail is due to call bots / ISP, no human talks that much

I still can't figure out the spike around 5500 texts

I think that I should note that I am not committing on the right / wrong of the utilization of this data, merely that I thought it was cool that something I learned in the classroom was actually useful.

*The AI going into the voice to text translators, and then data mining on that information.
**I apologize for the terrible graphics.