Customer churn prediction in a slow fashion e-commerce context : An analysis of the effect of static data in customer churn prediction

Sammanfattning: Survival analysis is a subfield of statistics where the goal is to analyse and model the data where the outcome is the time until the occurrence of an event of interest. Because of the intrinsic temporal nature of the analysis, the employment of more recently developed sequential models (Recurrent Neural Network (RNN) and Long Short Term Memory (LSTM)) has been paired with the use of dynamic temporal features, in contrast with the past reliance on static ones. Such an abrupt shift of policy has left open the challenge of understanding how those two kinds of features influence the predictive capabilities of models. This thesis aims at assessing the effect of combining static and dynamic features on the most commonly used models in survival analysis. In doing so, we compare the error measurements of such models with dataset composed of purely dynamic features or a combination of static and dynamic ones. Empirical measurements have shown that models respond differently to the addition of static features to the analysis, with more complex, sequential models like the LSTM struggling to deal with the added data complexity (with a 12% increase in error), while non sequential models see reductions of up to 14.7% in error. The thesis also includes a clusterization task aimed at aiding the interpretation of survival analysis outcomes.