Matt Stabeler

We explored a small portion of the Studivz dataset; around 4 months between Oct 2007 and Feb 2008 (illustrated below).

Studivz Weekly and Daily Activity, showing limited periods used, and message peaks (new year)

Because the dataset is so large, it was not practical to use all of the 28,000 nodes for testing, so, we took a small number of nodes (10) and travelled a number of hops (2) into their connected peers in a connected time graph, and used only this set of nodes to run out simulation. In this instance, we used 385 nodes. We again used MEAN,MEDIAN,80th and 20th percentiles in connected time ratios (described here) as thresholds on edge weights for Community finding using HGCE,KCLIQUE and LinkClustering. (Moses, Random and InfoMap are not thresholded).

Values used:

• Mean: 0.0001171252,
• Median: 5.66346e-05,
• 80th Percentile: 0.0001579604,
• 20th Percentile: 1.1558e-06

For InfoMap, this resulted in the graph below:

InfoMap Clustering of the reduced node set in Studivz. All edges shown.

KCLIQUE performed quite poorly on this for 80Pc, as there are some clear structures that it did not find.

KCLIQUE 80PC Threshold on Studivz Limited period/nodes

It performed slightly better on 20th percentile thresholding, but again a number of clear structures missed.

KCLIQUE 20PC Threshold on Studivz Limited period/nodes

LinkClustering for 80pc is shown below:

LinkClustering studivz dataset - 80pc

HGCE finds some structure to the network, and successfully finds hierarchy:

HGCE Community Strucutre for Studivz. Each node represents a community.

Result of Routing:

Studivz Dataset results of routing

For Prophet UF and H&W the results are below:

Studivz Delivery Ratio for unlimitedflood, hold and wait, and Prophet

It seems that even unlimited flood can only pentrate to 20% delivery ratio