I have an interesting problem in my company.
There is a product, each item of which, has a region-time specific code when it is released in the market.
Say, there are 15 features on the basis of which a code is created. The code is created by a lot of tiny rules (we don't know, they won't say). They have a pipeline where they predict (classify the code based on historical data dumps) this code given the product details and move along with some task.
Now, so far they dealt with 100 such code classes, 30 digit code is generated with the 15 features. And there are 100 such codes generated by the rules. So, their pipeline was fine and dandy.
Now, we are told to handle worldwide sales. Which means the code classes shoot off. Many more codes are now possible each with some changes due to region.
Earlier they used classification. And it was fine with 100 classes. But now classification with 10 K classes will probably won't work (does it even make sense?).
How to model the problem as a classification or regression problem with 10 K classes, the code is 15 digit alphanumeric thing?
Are they nuts in using machine learning (I think so) for this scenario, why not just find out what are the rules for code generation?