As the telecommunications market has reached the stagnation phase, it has become
cruicial for service providers to retain the existing subscribers. In order to
prevent the existing subscribers to switch their subscription to competition, i.e.
perform churn, the provider has to determine which subscribers will churn in the
near future and take appropriate action in order to prevent it.
The existing churn prediction models either consider subscribers as individuals
or additionaly take into account the data, related to users' interconnections.
In this research we have investigated the relevance of social network to churn.
Accordingly, based on variables that describe the subscribers in the context of
social network, we have built a churn prediction model and con_rmed the relevance
of social network with prediction results.
Due to large ammount of call connections data, the requirements of existing
prediction models surpass the available computer resources that service provider
would have to allocate to make predictions on real data. Therefore, due to large
ammount of data, the reduction of model complexity represents lower requirements
for computer resources which directly reects in lower _nancial expenses.
Due to these facts, we have proposed a model for churn prediction which is simpler
than the models proposed up to date.
The core of the thesis is a proposal of a simple model which predicts churn
upon taking into account previous churn among neighbors and their phone call
connections to the observed user. We have decided for such model in order to
clearly con_rm the assumption that the social network is highly relevant to churn
and it is possible to predict churn solely by observing social network parameters.
The simplycity is justi_ed with the comparison of the proposed models with
existing, more complex models. The comparison reveals that the proposed models
achieves comparable or better results than the existing, more complex models.
With purpose to make prediction in real time, we propose a model which is
based on _ndings related to the importance of social network. The results of the
model reveal that in case of using the proposed model to make churn predictions
on regular, daily or weekly basis, it is possible to capture considerable higher
percentage of subscribers who are going to cancel the subscription plan, compared
to random selection.