Prediction of mobile network subscriber satisfaction by using network probing experience measures and machine learning

The mobile telecommunications market is a highly competitive and mature market and mobile network operators (MNOs) increasingly rely on the quality and reliability of the core services they offer to distinguish themselves from other market players. Customer satisfaction plays a crucial role in such a landscape where negative word of mouth could severely damage the reputation of a business. Customer satisfaction has therefore become a key differentiator for many companies. A popular metric to track customers' experience with a business is the Net Promoter Score® (NPS). NPS is measured by customer surveys, prompting them to answer a simple question: “How likely are you to recommend company X to a friend or colleague?” The response ranges between zero, representing not likely, to ten, representing very likely. The score value is obtained by grouping responses into three categories: Promoters, Neutrals or Detractors, and calculating the percentage difference between promoters and detractors. The more positive the value, the better overall customer perception is likely to be. A key shortcoming of NPS is that it does not provide tangible and directly interpretable reasons for customer responses. This thesis aims to establish whether machine learning models, combined with network experience data collected by passive probing of mobile network interfaces, can accurately predict whether a subscriber will likely be a detractor. In addition, we would like to understand which network experience metrics are the best indicators of poor performance and negatively influence subscriber perception. We make use of survey and network data sourced from a large mobile network operator in South Africa over six months to create modelling features for cross validation of classification models with varying complexity to predict the NPS class of subscribers. We find that mobile network data provided by present Customer Experience Management (CEM) systems may not be ideal for use in machine learning applications. The standard library of metrics and data structures used to perform classical CEM requires much effort to clean and prepare it as viable input to machine learning models. In addition, we find that all tested machine learning models, whether linear or non-linear, are poor predictors of NPS. This suggests that NPS may instead be driven by other factors, such as pricing or the interaction of customers with other processes that are more important and not represented within the present data.

Grantor dc:publisher.institution

Department of Statistical Sciences

Year dc:date.issued

2025

Author dc:creator

• Kruger, Martin

Advisor dc:contributor.advisor

• Pienaar, Etienne

• Machine learning

Language dc:language.iso

en

Handle dc:identifier.uri

http://hdl.handle.net/11427/42363

OAI identifier oai:identifier

oai:open.uct.ac.za:11427/42363