New Scikit-Learn is More Suitable for Data Analysis
Around last year December, Scikit-Learn released a major stable update (v. 1.2.0–1) and finally I get to try some of the highlighted new features. It's now more compatible with Pandas and a few other features will also help us in regression as well as classification tasks. Below, I go through some of the new updates with examples of how to use them. Let's begin!
Compatibility with Pandas:
Applying some data standardization before using them for training an ML model like regression or neural net is a common technique to make sure different features with various ranges get equal importance (if or when necessary) for predictions. Scikit-Learn provides various pre-processing APIs like StandardScaler , MaxAbsScaler etc. With the newer version, it is possible to keep the dataframe format even after the pre-processing, let's see below:
from sklearn.datasets import load_wine
from sklearn.model_selection import train_test_split
########################
X, y = load_wine(as_frame=True, return_X_y=True)
# available from version >=0.23; as_frame
########################
X_train, X_test, y_train, y_test = train_test_split(X, y, stratify=y,
random_state=0)
X_train.head(3)
The newer version includes an option to keep this dataframe format even after the standardization:
############
# v1.2.0
############
from sklearn.preprocessing import StandardScaler
scaler = StandardScaler().set_output(transform="pandas")
## change here
scaler.fit(X_train)
X_test_scaled = scaler.transform(X_test)
X_test_scaled.head(3)
Before, it would have changed the format to a Numpy array:
###########
# v 0.24
###########
scaler.fit(X_train)
X_test_scaled = scaler.transform(X_test)
print (type(X_test_scaled))
>>> With the dataframe format remaining intact, we don't need to keep tabs on the columns, like we needed to do with the Numpy array format. Analysis and plotting become easier:
fig = plt.figure(figsize=(8, 5))
fig.add_subplot(121)
plt.scatter(X_test['proline'], X_test['hue'],
c=X_test['alcohol'], alpha=0.8, cmap='bwr')
clb = plt.colorbar()
plt.xlabel('Proline', fontsize=11)
plt.ylabel('Hue', fontsize=11)
fig.add_subplot(122)
plt.scatter(X_test_scaled['proline'], X_test_scaled['hue'],
c=X_test_scaled['alcohol'], alpha=0.8, cmap='bwr')
# pretty easy now in the newer version to see the effect
plt.xlabel('Proline (Standardized)', fontsize=11)
plt.ylabel('Hue (Standardized)', fontsize=11)
clb = plt.colorbar()
clb.ax.set_title('Alcohol', fontsize=8)
plt.tight_layout()
plt.show()
Even when we build a pipeline, each transformer in the pipeline can be configured to return dataframes as below:
from sklearn.pipeline import make_pipeline
from sklearn.svm import SVC
clf = make_pipeline(StandardScaler(), SVC())
clf.set_output(transform="pandas") # change here
svm_fit = clf.fit(X_train, y_train)
print (clf[:-1]) # StandardScaler
print ('check that set_output format indeed remains even after we build a pipleline: ', 'n')
X_test_transformed = clf[:-1].transform(X_test)
X_test_transformed.head(3)
Fetching DataSet is Faster and More Efficient:
OpenML is an open platform for sharing datasets and the Dataset API in Sklearn offers fetch_openml function to fetch data; With the updated Sklearn, this step is more efficient in memory and time.
from sklearn.datasets import fetch_openml
start_t = time.time()
X, y = fetch_openml("titanic", version=1, as_frame=True,
return_X_y=True, parser="pandas")
# # parser pandas is the addition in the version 1.2.0
X = X.select_dtypes(["number", "category"]).drop(columns=["body"])
print ('check types: ', type(X), 'n', X.head(3))
print ('check shapes: ', X.shape)
end_t = time.time()
print ('time taken: ', end_t-start_t)Using parser='pandas' makes a drastic improvement in runtime and memory consumption. One can easily check the memory consumption using psutil library as:
print(psutil.cpu_percent())Partial Dependency Plots: Categorical Features
Partial dependency plots existed before too, but only for numerical features, now this has been extended for categorical features.
As described in the Sklearn documentation:
Partial dependence plots show the dependence between the targets and a set of input feature(s) of interest, marginalizing over the values of all other input features (the ‘complement' features). Intuitively, we can interpret the partial dependence as the expected target response as a function of the input features of interest.
Using the ‘titanic' dataset from above, we can easily plot the partial dependence of categorical features:
With the code block above, we can get partial dependency plots as below:
With version 0.24, we would be getting a value error for categorical variables:
>>> ValueError: could not convert string to float: 'female'Directly Plot Residuals (Regression Models):
For analyzing the performance of a classification model, within Sklearn metrics API, plotting routines like PrecisionRecallDisplay , RocCurveDisplay existed in older versions (0.24); In the new update, it is possible to do similar for regression models. Let's see an example below:
While it's always possible to plot the fitted line and residuals using matplotlib or seaborn, after we've settled down with the best model, it's great to be able to quickly check the results directly within Sklearn environment.
There are a few more improvements/additions available in the new Sklearn, but I found these 4 major improvements to be particularly useful for standard data analysis more often than not.
References:
[1] Sklearn Release Highlights: V 1.2.0
[2] Sklearn Release Highlights: Video
[3]All the plots and codes: My GitHub
_If you're interested in further fundamental Machine Learning concepts and more, you can consider joining Medium using My Link. You won't pay anything extra but I'll get a tiny commission. Appreciate you all!!_
