REVISTA DYNA MANAGEMENT

Key words:

churn rate, inteligencia artificial, ensamble, estimador, modelo predictivo, oversampling, undersampling, datos desequilibrado, artificial intelligence, assembly, estimator, predictive model

Article type:

ARTICULO DE INVESTIGACION / RESEARCH ARTICLE

Section:

RESEARCH ARTICLES

ABSTRACT:
The study of the churn rate by companies, as a percentage of customers or subscribers who stop using their products or services during a certain period, is a common practice within a company. This study allows detecting behaviour patterns associated with whether the customer wants to stay or not. This detection of patterns can be done by using Machine Learning techniques related to the use of supervised learning models. This detection of abandonment will allow the company to address the retention strategies it considers relevant to avoid unnecessary economic losses. This paper presents the case study of the telecommunications company Orange, data from the SIDKDD 2009 competition.
In order to compensate the unbalanced nature of the data and, therefore, to promote the detection of clients leaving, two techniques are proposed: the use of the SMOTE algorithm and a customized version of a subsample assembly. Having analysed the results, it has been found that the training of the algorithms with a balanced dataset allows to improve considerably the capture of the customers leaving at expense of a certain penalty in the accuracy of the model. Finally, a range of estimate assemblies is incorporated so as it records every possible combination of the predictions of the predictive models used, in such a way that different balances are obtained between accuracy and detection of customer leaving.

Key Words: churn rate, artificial intelligence, assembly, estimator, predictive model, oversampling, undersampling, unbalanced data