Finding spatial outliers in collective mobility patterns coupled with social ties

ABSTRACT

Currently the increase in the variety and volume of data sources is demanding new data analytical workflows for exploring them concurrently, especially if the goal is to detect spatial outliers. In this paper, we propose a data analytical workflow for exploring Call Detail Records in conjunction with geotagged tweets. The aim was to investigate how massive data point observations can be analyzed to detect spatial outliers in collective mobility patterns that are coupled with social ties. This workflow consists of analytical tasks that are developed based on the a-priori assumption of two isometric spaces where Natural Language Processing techniques are used to find spatial clusters from geotagged tweets in a Social Space which are later used to aggregate the Call Detail Records generated by antennas located in the Mobility Space. The dynamic weighted centroids that are given by the mean location of the number of calls per hour of all antennas that belong to a particular cluster are used to compute Standard Deviation Ellipses. The longer the period of time a weighted centroid stays outside of the 99.7% probability region of an ellipse, the highest the likelihood that they are spatial outliers. The workflow was implemented for the city of Dakar in Senegal. The results indicate that the further the hourly weighted centroids are skewed from the normal mean of an ellipse, the stronger the influence of a cluster is in finding spatial outliers. Furthermore, the longer the period of time the outliers stays outside of the 99.7% probability region of an ellipse, the highest the likelihood that the outliers are genuine and can be associated to extraordinary events such as natural disasters and national holidays.