On the application of analytical techniques to mobile network CDRs for the characterisation and modelling of subscriber behaviour

Abstract

This thesis describes work completed on the characterisation and modelling of the behaviour of subscribers in mobile phone networks implementing dynamic pricing by means of Call Detail Records (CDR) datasets analysis. The initial focus of the work was the development of algorithms and visualisation techniques to facilitate an investigation of general subscriber behaviour in the Ugandan mobile phone network in which the CDR datasets was recorded. A cell site location estimation algorithm was developed which provided initial estimates for the location of cell sites which in 80% of all cell sites was within 20km of their true location. A joint calling and mobility analysis showed that the majority of subscribers visited fewer than six cells and that these subscribers usually made fewer than six calls per day. A coarse-graining strategy was also applied to identify the most significant communications corridors between the major urban centres in the country. The second component of this work was an investigation into what insights the CDR datasets could provide relating to aspects of social behaviour and economic activity in Uganda. A methodology for identifying centres of population concentration for residential and work activities was developed. This analysis highlighted a pattern where economic activity appeared to be concentrated around a small number of urban centres. Significant regional insularity in terms of population movement and communication was also observed along with regional behavioural homogeneity, particularly in the populations of Eastern and Western regions of the country compared to behaviour in the region around the capital and in the economically under-developed Northern region. The final component of this work focussed on the development of a novel Agent Based Model (ABM) for the simulation of subscriber behaviour based on patterns observed in the CDR datasets. Results showed that the simulated behaviour observed in the ABM bore strong similarities to that observed in the CDR datasets. The ABM was then utilised as a tool to evaluate the likely levels of revenue generation for a number of different dynamic pricing algorithms. Results obtained using the ABM suggested that all the dynamic pricing regimes investigated would likely result in revenue losses between 6% and 30% with a fixed subscriber base). Additional results also indicated that the scope for using optimisation techniques for revenue maximisation through real time control of dynamic pricing algorithms may be limited.