Output mode-specific parameters in __main__.model_stats

[slebastard]

For tracking purposes as well as to produce explainability analyses, it would be useful for each model run with python3 -m predictsignauxfaibles to output model-specific parameters into model_stats.

A use case would for the logistic regression would be to output its coefficient and baseline variables in order to find the contribution of each variables to the score a posteriori. Note that this could exist independently from the explanations pushed to the collection Scores:

• in the collection Scores, we could output the contribution of a group of variables for each SIRET input
• in model_stats, we would store model-specific parameters

A solution could be to specify in model_config.py the name of model attributes to be saved into model_stats.json, and to have a function towards the end of __main__.run(args) that collects those parameters into model_stats
WDYT?

[vviers]

For tracking purposes as well as to produce explainability analyses, it would be useful for each model run with python3 -m predictsignauxfaibles to output model-specific parameters into model_stats.

A use case would for the logistic regression would be to output its coefficient and baseline variables in order to find the contribution of each variables to the score a posteriori. Note that this could exist independently from the explanations pushed to the collection Scores:

• in the collection Scores, we could output the contribution of a group of variables for each SIRET input
• in model_stats, we would store model-specific parameters

I agree that this would be a nice to have

A solution could be to specify in model_config.py the name of model attributes to be saved into model_stats.json, and to have a function towards the end of __main__.run(args) that collects those parameters into model_stats

I can see how this implementation could add a lot of complexity to __main__ (which is already pretty big and hard to read)...
I think we should refactorize __main__ before we tackle this. For example by creating a SFModelRun class and implementing util functions or methods around it.

Let's talk about it soon 🙂