|Customer churn estimation with neural networks|
|Organisatsiooni nimi||Wise Ltd|
|Kokkuvõte||At companies such as Wise, which do not have subscription contracts but where customers transact whenever they see fit, it is quite hard to measure, or even define, customer churn. We can only say that when a customer has not used us for an unusually long time compared to their previous usage patterns, they have most likely churned - we can thus think of churn as a latent stochastic variable that we can only infer probabilistically. A classic way to treat churn in this fashion is the Buy Till You Die (BTYD) family of models, as explained eg in the “Counting Your Customers” the Easy Way: An Alternative to the Pareto/NBD Model" paper by Fader et al. These models are deliberately simple, which can be an advantage in many contexts, but limits their ability to represent complex customer groups and behaviors. |
We aim to integrate the BTYD approach in a neural network which predicts the next action by a given customer; this can allow us to investigate the factors driving churn in each particular case, as well as measure it more accurately. We already have the churn term implemented in the model (as a latent variable fitted via log-likelihood), and the code to generate synthetic data according to the BTYD model.
This project will tune the model with the focus on convergence of the loss term, and compare its performance with that of the BTYD model as implemented in the lifetimes Python package, on synthetic data, and if time allows, investigate that model's ability to quantify external factors' impact on churn (easy to inject into synthetic data).
|Lõputöö kaitsmise aasta||2021-2022|
|Juhendaja||Dr Egor Kraev|
|Nõuded kandideerijale||Familiarity with neural networks and log likelihood estimation, interest in customer analysis/marketing data science|