Original article was published on Artificial Intelligence on Medium
9 Changes to Note
1. Python 3.6 or newer only 🐍
Python 3.6 or newer is required to use scikit-learn 0.23. No more Python 3.5. It’s a good excuse to upgrade, if you need one. 🙂
2. Interactive pipeline graphics ⬇
You can now visualize your pipeline in an interactive graphic, right inside your notebook. The data flow from the top to the bottom. Technically, you’re visualizing a DAG (Directed Acyclic Graph). Here’s an example:
You can see the code that accompanies this article on GitHub here.
Just add the following code, make and fit a pipeline, and the graph appears! 😀
Pipelines are a great feature of scikit-learn! Pipelines and ColumnTransformers are powerful, but can be a tricky for newcomers to grasp. These diagrams can help folks learn and understand what’s happening faster. 👏
ColumnTransformer object allows different transformations to be applied to different features. I suggest you create them with the
make_column_transformer convenience method. Note that if you pass columns through unaltered with the
passthrough argument, that doesn’t show up on the DAG. ☝️
3. Poisson and gamma GLMs have arrived 🎉
The Poisson and Gamma generalized linear models can be imported with
linear_model.GammaRegressor, respectively. Now you shouldn’t need to leave scikit-learn for scipy.stats or statsmodels if you need this functionality.
Poisson regression is often appropriate for count data and gamma regression is often appropriate when predicting the time between two Poisson events. If you are are looking for more information on when to use a gamma GLM see this Cross Validated post.
fit() doesn’t show you everything 🚫
fit() method will not show all the attributes of an estimator when you call it. Only the arguments that you changed are shown.
To show all the attributes, as in earlier versions, run this code:
Alternatively, just call the
get_params() method on the estimator to see the parameters. 😀
5. n_features_in_ shows you how many features 🔢
Most estimators now expose the
n_features_in_ attribute to display how many features were passed to the
Note that in a pipeline with OneHotEncoder
n_features_in will show you how many features go in, not how many are fit to the final model. ☝️
6. Easier sample dataset loading 🧭
Most sample datasets can be loaded into a pandas DataFrame more easily. Just pass the argument
as_frame=True. Then the
.data attribute is a DataFrame. For example, here’s how you can load the diabetes dataset:
diabetes = load_diabetes(as_frame=True)
df_diabetes = diabetes.data
Loading datasets from scikit-learn used to be a bit of a pain. It’s easier now, but still not as easy as seaborn. Note that
load_boston() returns the Boston Housing dataset but doesn’t implement
as_frame yet. ☝️
7. Avoid type hinting errors ⚠️
Scikit-learn now works with mypy without erring. If you’re using type hinting, this is nice. 😀
8. Improvements to experimental classes 🧪
HistGradientBoostingClassifier, the two LightGBM-inspired tree ensemble algorithms, are still experimental. They still need to be specially imported. However, they received a number of improvements. Same with
IterativeImputer — it’s still experimental and has been improved.
9. Plays nicer with new pandas dtype 🐼
Speaking of scikit-learn imputers, they now accept the pandas nullable integer dtype with missing values — see my article on what’s new in Pandas 1.0 to learn about those. The continued friction reduction between pandas and scikit-learn is music to my ears.
You’ve seen the 9 most important changes in scikit-learn version 0.23.0. Now you can impress your friends and colleagues with your knowledge. 😉
I hope you found this guide to be helpful. If you did, please share it on your favorite social media so other folks can find it, too. 😀
I write about Python, SQL, Docker, data science, and other tech topics. If any of that’s of interest to you, follow me and read more here. If you want to stay up on the latest data science tips and tools, subscribe to my Data Awesome mailing list.
Happy sklearning! 🚀