Feature types and explainability

[Yasslight90]

Hi there,

I have three questions:

1- What feature types does FLAML support? I am sure it supports at least numerical and categorical features. Any other types?
2- Is it possible to force the type for a given feature?
3- Does FLAML provide an explanation for predictions? I am thinking of shap values for example.

Thank you so much

Yassine - Data scientist

[Chi Wang (sonichi)]

1. numerical, categorical and datetime.
2. if you pass a pandas dataframe, the type is specified for each column.
3. explanation is not built-in from flaml. Is there any package you are using for explanation now?

[Yasslight90]

Thank you for the quick answer!

3- No. I am evaluating the autoML-ness of many solutions and regarding their explainability I found out that they all use SHAP values to explain predictions.

[Chi Wang (sonichi)]

Thanks for the discussion. I'd like to get your opinion on: should explanation of predictions be part of an automl library? If so, what is the right thing to be automated?

[Yasslight90]

Hello,

Yes. From my point of view, autoML solutions should definetely be able to explain their outputs, especially when used in critical applications: Today, most of the existing solutions (Google autoML tables, Prevision.io, H2O, etc) use shapley values which are not very trustworthy.

Moreover, what I think should be provided to the user is: 1) a global interpretation (for example, by providing him/her with a feature importance plot) and 2) a local interpretation (by computing the importance of each feature in every prediction).

Cheers

Yassine

[Chi Wang (sonichi)]

Thanks for the input. It'll be nice to have a real use case where this feature is requested. If you have one, I'll be happy to hear it. Feel free to chat on gitter if you like.

[Yasslight90]

a simple use case would be: a user asking for a bank loan. Imagine that the bank relies on the predictions of a automated machine learning model to accept or not a bank loan. Clients who got a refusal would most likely want to understand this decision and the bank owe them an explanation; here we are talking about a local explanation. Also, bank advisor would want to know what are the most important variables in the classification (acceptance, refusal); here we are talking about a global interpretation.

[Kevin (kevball2)]

Are you looking for something like this? https://github.com/microsoft/responsible-ai-toolbox

Azure AutoML has some of these features in preview
https://docs.microsoft.com/en-us/azure/machine-learning/how-to-machine-learning-fairness-aml

[Yasslight90]

Thanks. However, I am not looking for a solution for explainability. I am just highliting the importance of having it integrated into the autoML process :-). For example, you would integrate Fairlearn into FLAML ...

[Qingyun Wu (qingyun-wu)]

Hi Yasslight90 and Kevin (@kevball2),

Thank you for your insightful comments. There are several related TODOs included in our roadmap:

• Integrate Fair AutoML, which leverages FairLearn to improve the fairness of the resulting ML models from FLAML. This is ongoing research. We plan to do the integration in the near future. Feel free to check out our paper and share your comments.
• Prediction quality. I believe this also contributes to improving explanation. The idea is to report the uncertainty of a model's predction in addition to the accuracy, e.g., via conformal prediction.
• Visualizing feature importance, explanation, and optimization history. This is still in an early stage investigating what should be included. Here are my understandings:
  - Lime can provide local explanations (e.g., feature importance for specific instances) as shown in this example.
  - Shap can provide both local and global explanations via shap values.
  And for integrating them into the AutoML process: a naive integration I can think of is just to apply these explainers (and the
  confidence level) at the end of the AutoML process to the resulting model (the best model) of FLAML. Do you have additional comments on where this explanation should happen in the AutoML process?

Thank you!