Check out the new USENIX Web site. next up previous
Next: Effects of Similarity Thresholds Up: Results Previous: Effectiveness of the Correlation


Configuration

The algorithms use two internal constants that we believe are not sensitive to particular network connection patterns. The group formation phase of the role classification algorithm (see Section 4.1) requires a constant $0 \leq \alpha \leq 1$ to keep a host h from forming groups with other hosts that have less than a fraction $\alpha$ of the number of connections that h has. The group merging phase keeps the two groups from merging if the average number of connections of a group is not within $0 \leq \beta
\leq 1$ of the other's (see Figure 3).

We set $\alpha = 0.6$ and $\beta = 0.5$. Our experiments with both Mazu and BigCompany networks indicate that the default values work well on at least two rather different networks. We believe that, in general, it will not be necessary to adjust these constants. Nevertheless, we plan to expose these parameters to network administrators so that they can adjust them along with the similarity thresholds to achieve grouping results that most reflect their intuition of the network structure.

Figure 6: Number of Groups vs. Slo

Figure 7: Number of Groups vs. Khi


next up previous
Next: Effects of Similarity Thresholds Up: Results Previous: Effectiveness of the Correlation
Godfrey Tan 2003-04-01