Automating Clustering Analysis of Ivory Coast Mobile Phone Data: Deriving Decision Support Models for Community Detection and Sensemaking

摘要

Sensemaking involves numerous levels of processing and logic in order to achieve automated decision support. Many of these concepts derive from the realm of pattern recognition. The data under consideration frequently is observed in a noisy environment and so one of the first steps involves preprocessing the data to suppress noise and isolate the data signal. Patterns within the data are often used to improve signal detection and aid identification of the data in the quest to produce actionable information. A critical step of making sense from raw or partially processed data and other aspects of decision support is to organize information, which frequently involves grouping, partitioning, or clustering objects. However, there is typically an assumption that structure exists within the data, and the number of clusters is a required parameter for many of the clustering algorithms. A common approach to determine the best number of clusters is to iterate across a set of potential values the for number of clusters and evaluate the quality of the resulting clusters using some metric. In this paper, we present an automated approach to detect structure and improve automation of clustering algorithm parameters. We apply our approach to analyze a complex, dynamic multiple edge set network that was used to model call data from the Ivory Coast compiled from France Telecom/Orange anonymized call records over a 5 month period.