Sensitivity of social network analysis metrics to observation noise

Social network analysis (SNA) is a methodology that represents interpersonal communications as directed graphs. SNA uses graph metrics to quantify different aspects of a group\´s communication patterns. This work supports our goal of identifying terrorist activity based on the atypical SNA metric values of their communication patterns. Imperfect observation is given, so it is necessary to understand how various SNA metrics react to observation error. In this paper, we analyze this sensitivity, in order to guide decisions about which metrics should be used in terrorist activity detection. Results have shown, for example, that the radius metric is extremely sensitive to imperfect observation - exhibiting up to 7000% error at 70% observability. Characteristic path length is much less sensitive to observability (often less than 20% error at 70% observability), and the error in density is moderately well approximated as a Gaussian whose parameters are linear functions of observability. Our data is computed on random synthetic "phi" social interaction graphs, as presented in Duncan Watts\´ book "Small Worlds".