< Terug naar vorige pagina

Publicatie

Leveraging Mobile Phone Data and Social Network Analytics for Profit Driven Modeling.

Boek - Dissertatie

In the rapidly growing world of data science and analytics, data has become an asset that gives companies a competitive edge in fierce and saturated markets. To make the most of all the available data and resources, it is necessary to be mindful of each phase in the analytics process and especially by using the appropriate data with the right techniques and an evaluation that is befitting for the problem at hand. This dissertation considers the added value of combining mobile phone data with social network analytics and the appropriate evaluation measures with the goal of improving procedures in the analytics process and facilitating better usage of data science for decision making in two fields of industry, namely churn prediction and credit scoring.Alternative data sources can be challenging to incorporate in analytics applications because they tend to be unstructured and difficult to process. On the other hand, they may possess valuable information that should not be overlooked. This disseration utilizes mobile phone data, which has the benefit of providing a true representation of people's behavior and can thus be used in various ways to provide different types of information depending on the approach.Humans are social beings and their actions are influenced and controlled by those around them. Consequenlty, social networks are important in day to day life. When making predictions involving human behavior, it is reasonable to take into account the social component and influence from others. Preprocessing the mobile phone data in a specific way to construct networks gives the opportunity to extract alternative information that provides a novel perspective on the observations in the dataset and facilitates new applications of social networks. The two applications of analytics this dissertation focuses on, are meant to be deployed in industry and as such should be evaluated appropriately. In a business context, this means focusing on profitability and economic impact. Therefore, we apply a novel performance measure that is based on profit and has been customized for churn prediction as well as credit scoring. This offers the industries the opportunity and incentive to utilize the abundance of customer data they have to increase the impact of their data-driven decisions.This thesis contains four projects where mobile phone data and social networks are used to enhance the impact of analytics models from a profit driven perspective. The first project presents a benchmarking study of social network techniques for churn prediction in the telecommunication industry. The study provides a ranking of two dozen network learning techniques based on their predictive performance, as well as a comparison of churn prediction models applied to features and influence scores extracted from the networks. Furthermore, various network architectures are explored in order to identify which one facilitates the most accurate predictions. In the second project, mobile phone data and social networks are used to predict the creditworthiness of potential borrowers. The study combines mobile phone data with bank history data of over a million bank customers to build social networks that are featurized before training predictive models. The models are subsequently evaluated using both statistical and economic performance measures which show that features representing personal calling behavior are most predictive of creditworthiness. The third project approaches the networks from a dynamic perspective, where a sequence of networks is built to provide a time series of behavior based features that are classified using similarity forests. This novel technique of building and featurizing networks performs better when predicting churn in the long term. In the fourth project, the expected maximum profit measure for churn prediction is extended to take into account the variability of customer lifetime values, thus giving a more accurate representation.In summary, this dissertation presents how the combination of alternative data, social network applications and evaluation measures can increase the impact of the analytics process and facilitate better usage of data science for decision making in industry. Thereby, it takes into account the natural behaviour of people in order to enhance the customer journey and improve relationships with the organizations that are an inevitable part of people's lives.
Jaar van publicatie:2018
Toegankelijkheid:Open