I have a large set of simulation logs for a market simulation of which I want to learn from. The market includes:

  • customers
  • products (subscriptions)

The customers choose products and then stick with them until they decide on a different one. Examples could be phone, electricity or insurance contracts.

For every simulation I get the data about the customers (some classes and metadata) and then for each round I get signups/withdrawals and charges for the use of the service.

I am trying to learn a few things

  • competitiveness of an offering (in relation to the environment/competition)
  • usage patterns of customers (the underlying model is a statistical simulation) depending on their chosen tariff, time of day and their metadata + historical usage
  • ability to forecast customer numbers for each product

The use cases are all very applicable to real world data although my case is all a (rather large) simulation.

My problem is this: What kind of learning is this? Supervised? Unsupervised? I have come up with various hypotheses and cannot find a definite answer for either.

  • Pro Supervised: For the usage patterns of the customers I have historical data of actual usage so I can do something similar to time-series forecasting. However, I don’t want to forecast simply off of their previous usage but also off of their metadata and their tariff choice (so also metadata in a way).
  • Pro Unsupervised: The forecasting of the “competitiveness” of a randomly chosen product configuration is hard to label even with historical data. The exact reason why a product has performed in a certain way is very high-dimensional. I do get subscription records about every product for every time slot though, so I guess some “feedback” could be generated. This might also be a RL problem though?

So obviously I need help pulling these different concepts apart so as to map them on this kind of problem which is not the classical “dog or cat” problem or the classical “historical data here, please forecast” timeseries issue. It’s also not a “learn how to walk” reinforcement problem as it’s based on historical data. The end goal is however to write an agent that generates these products and competes in the market so that will be a reinforcement problem.



You must log in to answer this question.

Browse other questions tagged .