Churn Prediction using hierarchical generative models

Predicting customer churn creates opportunities to target customers with a marketing action or a promotion to prevent them from leaving. In this paper, a hierarchical generative approach will be applied in the context of churn prediction. The dependent variable and the covariates are modeled jointly conditioned on a deep latent structure, which resembles the hidden structure in neural networks. The conditional latent structure is capable of handling missing data and combining heterogeneous data types. Latent structures with multiple layers are non-linear and can model complex dependencies between the independent variables and the risk to churn. The hierarchical generative approach makes use of deep exponential families (DEFs). This class of models is able to extract a hierarchy of dependencies between latent variables. Similar to deep unsupervised feature learning, this analysis can improve predictions and provide extra insights into the nature of the data. The hidden layers in the DEFs enable the exploration of interesting structures in datasets. These patterns could help sales representatives in classifying customers according to their risk of churn, so that companies or managers can take more well-informed decisions