tau-mlwell/marginal-contribution-feature-importance

source code for Marginal Contribution Importance (MCI) method

marginal-contribution-feature-importance's Issues

Support for non-sklearn estimators - Support for already trained estimators

Is there support for generic estimators, i.e. any deep learning model not supported by sklearn and that do not possess that API? (e.g. pytorch, tensorflow).

How would one proceed if they wish to compute MCI on a simple neural network (that's already trained), for example? Let's say it has a model.predict(X) interface that allows inference.

If cross-validation is needed to compute the MCI score, then let's say I already have trained 10 models using exclusive kfold validation splits.

Report an error in mci/estimators/contribution_tracker.py

When do json.dump(state, f) in lines 54, a typeerror was reported: Object of type int64 is not JSON serializable. After debugging the data in the dictionary, it is found that the value in the dictionary corresponds to the numpy.int64 type. Normally, the value should be numpy.float64. Hope you can modify it