ensembl | The Ensembl Core Perl API and SQL schema | REST library

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.

We may loop over the list and apply the bitr

I think using a multidimensional array is a very good idea in this case.

First, you can get the simulations of example_sim() much cheaper using mapply(). Here an example with time=10 and npops=3. Use the same set.seed(42) and parameters and check for yourself.

I use much smaller parameters here so that you can easily check the result in your head.

xgboost has min_child_weight, but outside of the ordinary regression task that is indeed different from minimum samples. I couldn't say why the additional parameter isn't included. Note though that in binary classification, the logloss hessian is p(1-p) and is between 0 and 1/4, with values near zero for the very confident predictions; so in effect setting min_child_weight is requiring many currently-uncertain rows in each leaf, which may be close enough to (or better than!) setting a minimum number of rows.

It depends on the version of sklearn you are using. In versions past 1.0, models have a feature_names attribute when trained with dataframes that integrates the column names. There was a bug in this version that threw an error when training with dataframes. https://github.com/scikit-learn/scikit-learn/issues/21577

I'm not up to date with the new best practices for this yet, so I cannot say definitively how it should be set up. But I just side stepped the issue in my code for now. To get around this, I convert my dataframes to a numpy array before training

• Point is that you're using predict() rather than predict_proba()/decision_function() to define your y_hat. This means - considering that the threshold vector is defined by the number of distinct values in y_hat (see here for reference), that you'll have few thresholds per class only on which tpr and fpr are computed (which in turn implies that your curves are evaluated at few points only).

• Indeed, consider what the doc says to pass to y_scores in roc_curve(), either prob estimates or decision values. In the example from sklearn, decision values are used to compute the scores. Given that you're considering a RandomForestClassifier(), considering probability estimates in your y_hat should be the way to go.

• What's the point then of label-binarizing the output? The standard definition for ROC is in terms of binary classification. To pass to a multiclass problem, you have to convert your problem into binary by using OneVsAll approach, so that you'll have n_class number of ROC curves. (Observe, indeed, that as SVC() handles multiclass problems in a OvO fashion by default, in the example they had to force to use OvA by applying OneVsRestClassifier constructor; with a RandomForestClassifier you don't have such problem as that's inherently multiclass, see here for reference). In these terms, once you switch to predict_proba() you'll see there's no much sense in label binarizing predictions.

Apparently, scikit-learn TransformedTargetRegressor objects don't allow you to change the regressor used to predict, unless you re-fit the dataset on the new regressor in set_params. If you do this:

argsort "returns the indices that would sort an array," so here sorted_idx contains the feature indices in order of least to most important. Since you just want the 3 most important features, take only the last 3 indices:

Unfortunately, there is an issue with scikit-learn's SimpleImputer when it tries to impute string variables. Here is a open issue about it on their github page.

To get around this, I'd recommend splitting up your pipeline into two steps. One for just the replacement of null values and 2) the rest, something like this: