Optimal Sensing of Community Structure From Device Data - An Active Learning Approach
09 October 2015
As mobile communication devices have grown popular in the possession of the average individual, location information retrieved from such devices is becoming a promising mean to detect and analyze social behavior and structure. Furthermore, the use of high resolution in-door location or proximity data allows detecting very delicate and high-resolution social and organizational patterns. However, several challenges should be tackled when inferring social patterns from in-door location data of devices. For example, co-location is typically more noisy due to the nature of the signal collected under highly dynamic wireless networks (e.g., WiFi), and casual proximities that may occur due to the functionality of various popularly-shared building areas (e.g., cafeteria or classroom). These settings may obscure the detection of meaningful and prominent social structures. We present in this paper a highly accurate algorithm that allows to detect communities in such challenging scenarios and we augment it with an 'active learner'. The active learning enhances the algorithm's performance by sequentially choosing an optimal training set to improve semi- and unsupervised learning of social structure. We demonstrate the efficiency and accuracy of our approach on synthetic and real data sets in which we recover the underlying social or organizational structure from device location data.
