Classification Problematics : Feature Number Variance & Feature Repetition

Question

I have a harsh case study (in my mind).

The problem is I need make binary classification on Quality of Service (good or bad). I have a feedback on quality on groups of devices belonging to company. I have some good information on each device (same features for all devices). However, the number of devices can change from a company to another.

I would like to classify the Quality thanks to each devices information.

We usually aggregate the data to each company, but I want to try a new approach in order to be able to identify the device(s) in default.

I think to use ensemble method (calculating a score for each device and finally calculating global score for quality) or interaction_cst (HistGradientBoostingClassifier of Sklearn for instance). Nevertheless, the fact that the number of devices can vary is an obstacle for me. Also, because there are the same features for each devices, I think that using the same model to score if preferable but I haven't the quality of the device itself.

So, I you have advice on these problematics, I'll be glad to exchange with you :)

Answer 1

Given your problem, you can consider the following approaches:

Ensemble Method: As you mentioned, you can use an ensemble method to calculate a score for each device and then calculate a global score for quality. This approach can work well if you have a good model for scoring individual devices. However, it might not work well if the number of devices varies significantly between companies, as the global score might be biased towards companies with more devices.

Hierarchical Modeling: Another approach you can consider is hierarchical modeling, where you have a model for each device and a higher-level model for the company. This approach can handle varying numbers of devices between companies, as the company-level model can take into account the number of devices. However, this approach can be more complex and computationally intensive than the ensemble method.

Remember, It's often a good idea to try several approaches and see which one works best.