The impact of graph mining features on churn prediction in the telecommunication industry

Details

Supervisors

Lebichot, Bertrand ; Courtain, Sylvain

Faculty

Louvain School of Management

Degree label

Master [120] : ingénieur de gestion, à finalité spécialisée

Abstract

Recent studies proved that the use of social network analytics is successfully used in the telecommunication industry for churn prediction. This master thesis aims to know if there is a valuable process to implement graph mining features in actual Orange's churn prediction model. To do so, we need to collect Call Detail Records (CDR) from which we will be able to extract graph mining features. We both constructed supervised and unsupervised graph mining features such as PageRank and the eigenvectors. Supervised graph mining features are made up of relational learners which are composed of relational learners and optionally collective inference methods. We compared two model types using various evaluation metrics. The first model contains the actual Orange's churn prediction model, named the baseline. The second model is made up of features present in the baseline and the graph mining features that we previously created and described. As a result of this master thesis, we conclude that there is no added value to implement graph mining features in Orange's churn prediction model.