CLF | Academy / ASC Common LUT Format Sample Implementations | Learning library

You can use list of dict of parameter combinations instead of a dict.

For example if you want to tune C, penalty, and solver by separating the solvers to different combination, you can do it by this way:

You could call the load_iris function without any parameters, this way the return of the function will be a Bunch object (dictionary-like object) with some attributes. The most relevant, for your use case, would be bunch.data (feature matrix), bunch.target and bunch.feature_names.

This library is not well-documented, and apart from the axis and the general color for both line and area, there seems to be nothing that you can pass on to the plot. But we can retrieve the plotted objects (in this case, the plotted Bezier curve consists of a Line2D and a PathPatch object) and modify them:

Now I checked again and I noticed that I put the function to wrong place. I will answer it below.

The reason why the edge weights are plotted is that the weights are not updated after removing a node. Hence, pos and labels in your script should be recalculated after removing the node:

In your example, you're using a DecisionTreeClassifier which by default support targets of dimension (n, m) where m > 1.

However if you want to have as result the marginal probability of each class then use the OneVsRestClassifier.

Notice that CalibratedClassifierCV expects target to be 1d so the "trick" is to extend it to support Multilabel Classification with MultiOutputClassifier.

Full Example

Not very sure how you trained your model or how the preprocessing was done. The leaf dataset has about 100 labels (species) so you have to take care to split your test and train to ensure an even split of your samples. One reason for the weird accuracy could be that your samples are split unevenly.

Also you would need to scale your features:

[This is a partial answer only, unfortunately.]

This is a Windows problem. I tried reproducing your results on macOS, and found a different, interesting behavior. I modified your code to distinguish between the C fwrite and the C++ std::fwrite, and I added code to write using the lower-level Posix write.

This is the C++ code:

As you would already be aware that the whole idea of feature importances is bit tricky for the case of LogisticRegression. You can read more about it from these posts:

1. How to find the importance of the features for a logistic regression model?
2. Feature Importance in Logistic Regression for Machine Learning Interpretability
3. How to Calculate Feature Importance With Python
   

I personally found these and other similar posts inconclusive so I am going to avoid this part in my answer and address your main question about feature splitting and aggregating the feature importances (assuming they are available for the split features) using a RandomForestClassifier. I am also assuming that the importance of a parent feature is sum total of that of the child features.

Under these assumptions, we can use the below code to have the importances of the original features. I am using the Palmer Archipelago (Antarctica) penguin data for the illustration.