Enhancement of explainability

[technocreep]

Shortly speaking, we'd like to know why models do certain prediction. For that sake we already implemented so called Point Explained. This tool considers any model as a black box and its explanation functionality is based on time series points/intervals perturbations. Relatively comprehensive description is provided in this notebook 📓

We wonder if there are more methods which could be implemented in Fedot.Industrial in order to develop understanding of model behaviour.

So this task would involve three subtasks:

• Dive deep into review paper, maybe this one, this looks interesting, and this one. One can search for papers by him/her self. What we need is to find out what methods are applicable in our case. So feel free to make yourself a cup of tea, read them carefully and discuss your findings with mentor 🍵
• Check whether SHAP or LIME approaches are good choises
• This is not necessary, but would be – to refactor PointExplained (only if it is needed)
• Implement your ideas
• Cover you code as much as possible with unit-tests
• Prepare notebooks with examples (such as one mentioned before)
• Feel exhausted (not really necessary) and happy (highly demandable)

Probably the module should be called like this:

industrial = FedotIndustrial(config)
industrial.fit(train_data)
industrial.predict(test_data)

industrial.explain(some_params) –> vis, plots, diagrams