Timestamp-based correlation measures for finding hidden groups in chat rooms

Abstract

This thesis describes a new two-step algorithm for finding hidden groups from chat transcripts, that is, transcripts of communication where the recipient of a message is not known. The algorithm is presented in two steps: calculating a correlation value between every pair of users in the chat transcript, and finding clusters in the weighted undirected graph that results. The inter-user correlation can be calculated in a number of different ways, some of which are accomplished by projecting individual user transcripts into an inner product space. The clustering step uses the existing iterative-scan algorithm, with some new modifications. This approach is found to work under limited conditions.