Community Detection on Proximity Networks

Bluetooth or Wi-Fi access point connections allow us to discover people’s physical proximity. However, their analysis is often difficult because the collected data is noisy and misleading. In this study, we use network modeling to analyze proximity information from these types of data sets. We extracted proximity networks from three different systems: Haggle Infocomm conference and MIT Reality Mining Bluetooth connections; and Sabanci University Wi-Fi access point connections, such that it is extracted as a static proximity network for the first time in this study. We explored both the extracted networks’ properties and the 12 community detection algorithms’ results. According to the descriptive analysis and statistical tests of all the setups, the performance of the algorithms depends on the data sets. The Haggle Infocomm network is noisy enough to make community detection difficult. However, the other two networks are suitable for the analysis. They contain a distinguishable community structure. Partitioning algorithms find large communities, while overlapping algorithms can detect smaller ones. The EMOC, which we proposed in our previous work, is able to find tiny communities with more overlapping nodes than any other algorithm. Such communities may correspond to small groups of friends in the same place in proximity networks.