Matt Stabeler

After a few iterations of bug fixes in the Bubble H code, it finally gave some sensible results. Shown below for MIT-OCT and MIT-NOV, where the community finding was done using a training set, and the test was done on the test set. Results are compared with the previous CFAs and also the  routing schemes we originally compared to (Bubble, PBR, Prophet, Unlimited Flood). GCEH was run multiple times with varying parameters, and the output chosen to drive BubbleH was chosen by picking a good looking result, this is a flaw in the process.

For MIT-OCT, the data in mit-oct-training.gce_output_K4_ST0.5_MAP0.5_e0.25_0.2.dat was used, which looks like:

  0(101)
  |
   - 1(71)
  |
   - 4(40)
  |  |
  |   - 2(16)
  |  |
  |   - 3(20)
  |  |
  |   - 5(28)
  |
   - 6(63)
  |  |
  |   - 7(47)
  |  |  |
  |  |   - 8(24)
  |
   - 9(66)

For MIT-NOV the data in mit-nov-training.gce_output_K3_ST0.5_MAP0.5_e0.15_0.2.dat was used, which looks like:

  0(101)
  |
   - 1(39)
  |  |
  |   - 2(16)
  |  |
  |   - 3(22)
  |  |  |
  |  |   - 4(14)
  |  |
  |   - 5(8)
  |
   - 6(30)
  |
   - 7(24)
  |
   - 8(63)

MIT-OCT

MIT-NOV

It is clear that MIT-NOV seems to have a better delivery ratio overall, this is probably due to the increased activity in December (MIT-NOV-TEST), compared to that in November (MIT-OCT-TEST), as seen in the activity plot below.

The average number of connections (bluetooth contacts) per week in the MIT Dataset

A thorough investigation would mean running the output of the many parameters used for the GCEH algorithm, and running the simulation over the whole lot – complicated, but possible.

UPDATE: Using MIT-NOV-CHEAT dataset – i.e. allowing the use of data from the test period, gives a much better result – see below.

MIT-NOV-CHEAT

The candidate hierarchy was taken from similar parameters as previously: edge_list.dat.gce_output_K-3_ST-0.5_MAP-0.5_E-0.15_Z-0.2.dat which looked like:

  2(77)
  |
   - 0(52)
  |  |
  |   - 1(18)
  3(79)

This was slightly different from the previous hierarchies, in that it has less communities, but it demonstrated an accurate looking community structure – i.e. 3 communities. The results indicate that BubbleH does better here than previous attempts, which is encouraging. I also tried attempted to run the KCLIQUE version of BubbleRAP, but, as discussed before, it cannot find any communities in the training+test period.