Structural mining of large-scale behavioral data from the internet

As the Internet becomes ever more pervasive in the lives of hundreds of millions of people, our understanding of its physical structure has outpaced our understanding of the dynamic patterns of traffic generated by its users. This work aims to develop a better understanding of the structure of Internet traffic in a manner consistent with individual privacy and computational constraints. I first examine network flow data from the Internet2 network, using it to form “behavioral networks” based on the flows attributable to specific network applications. The heavy-tailed distributions associated with these networks suggest unbounded variance and poorly defined means in distributions of user behavior. However, a novel application of hierarchical clustering to similarity data derived from these networks makes it possible to classify network applications robustly based on their observed behavior. I then focus on Web traffic, using a large collection of HTTP request data to build a weighted subset of the Web graph. Analysis of this weighted graph reveals more heavy-tailed distributions and the presence of a large body of stationary traffic. The traffic data are also shown to contradict key assumptions of the random surfer model used by PageRank. I conclude with the development of ABC, an behaviorally plausible agent-based model of Web traffic that incorporates backtracking, bookmarks, and a sense of topical locality. The ABC model is shown to approximate real user activity more accurately than PageRank on both artificial and empirically generated graphs.