| dc.description.abstract | The telecommunication’s market has changed to a saturated one during the past
few years. This makes it relatively costly to acquire new customers than to retain
existing ones. Recent research has further showed that identifying potential customers
who intend to leave a service provider for another (churners) and offering
them an incentive to keep their subscriptions can produce significant savings for
the telecommunication company. However, in most cases, it turned out that most
of the used techniques to solve this problem fails to address the complex relationship
between customer features and churn. Therefore, they struggle to achieve
acceptable prediction accuracy. This might be due to the fact that real customers’
data are large, unbalanced, with missing values, and with non-linear dependencies
between features. In this thesis, I investigated a new class of techniques for
predicting customer churn based on Gaussian processes - a recent approach to regression
and classification based on Bayesian reasoning principles. I developed
a Gaussian process based model for churn prediction and tested it on a real customers’
data belonging to a French telecommunication’s company (Orange). The
method obtained an area under the curve (AUC) score of 0.77. This score outperformed
most of the previously used techniques that I evaluated on the same
data set. The best of them had an AUC score of 0.73. To conclude, my empirical
results proved that Gaussian processes can provide state-of-the-art performance in
churn prediction, and the approach and results that I report offer a solid basis for
further development of Gaussian processes for churn prediction. | no_NO |
