Problem definition: Predict user's next event date, based on previous event occurrences. The aim is to inter-corporate time dependent and time independent features.
Data: +10 year transactional data generated by millions of users. 80% of users have less than 3 events.
Prediction: Next event date
I've gone through many similar questions, the most related ones were the following:
I am still unsure which method I should follow as the prediction should be on user level, but the users behaviour is really dissimilar. Shall I try to cluster the users first, and create different models for each cluster? Or am I better off fitting a distribution pattern for each user? If i use HNN, how can I intercorporate seasonality variables?
One of my biggest headache is how to transform the transactional dataset into a dataset which can be used for modeling. Previously, I've never dealt with a dataset with different entities (users) and also stochastic process.