Sample pipeline for text feature extraction and evaluation

The dataset used in this example is The 20 newsgroups text dataset which will be automatically downloaded, cached and reused for the document classification example.

In this example, we tune the hyperparameters of a particular classifier using a RandomizedSearchCV. For a demo on the performance of some other classifiers, see the Classification of text documents using sparse features notebook.

# Author: Olivier Grisel <>
#         Peter Prettenhofer <>
#         Mathieu Blondel <>
#         Arturo Amor <>
# License: BSD 3 clause

Data loading

We load two categories from the training set. You can adjust the number of categories by adding their names to the list or setting categories=None when calling the dataset loader fetch_20newsgroups to get the 20 of them.

from sklearn.datasets import fetch_20newsgroups

categories = [

data_train = fetch_20newsgroups(
    remove=("headers", "footers", "quotes"),

data_test = fetch_20newsgroups(
    remove=("headers", "footers", "quotes"),

print(f"Loading 20 newsgroups dataset for {len(data_train.target_names)} categories:")
print(f"{len(} documents")
Loading 20 newsgroups dataset for 2 categories:
['alt.atheism', 'talk.religion.misc']
857 documents

Pipeline with hyperparameter tuning

We define a pipeline combining a text feature vectorizer with a simple classifier yet effective for text classification.

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.naive_bayes import ComplementNB
from sklearn.pipeline import Pipeline

pipeline = Pipeline(
        ("vect", TfidfVectorizer()),
        ("clf", ComplementNB()),
Pipeline(steps=[('vect', TfidfVectorizer()), ('clf', ComplementNB())])
In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with

We define a grid of hyperparameters to be explored by the RandomizedSearchCV. Using a GridSearchCV instead would explore all the possible combinations on the grid, which can be costly to compute, whereas the parameter n_iter of the RandomizedSearchCV controls the number of different random combination that are evaluated. Notice that setting n_iter larger than the number of possible combinations in a grid would lead to repeating already-explored combinations. We search for the best parameter combination for both the feature extraction (vect__) and the classifier (clf__).

import numpy as np

parameter_grid = {
    "vect__max_df": (0.2, 0.4, 0.6, 0.8, 1.0),
    "vect__min_df": (1, 3, 5, 10),
    "vect__ngram_range": ((1, 1), (1, 2)),  # unigrams or bigrams
    "vect__norm": ("l1", "l2"),
    "clf__alpha": np.logspace(-6, 6, 13),

In this case n_iter=40 is not an exhaustive search of the hyperparameters’ grid. In practice it would be interesting to increase the parameter n_iter to get a more informative analysis. As a consequence, the computional time increases. We can reduce it by taking advantage of the parallelisation over the parameter combinations evaluation by increasing the number of CPUs used via the parameter n_jobs.

from pprint import pprint

from sklearn.model_selection import RandomizedSearchCV

random_search = RandomizedSearchCV(

print("Performing grid search...")
print("Hyperparameters to be evaluated:")
Performing grid search...
Hyperparameters to be evaluated:
{'clf__alpha': array([1.e-06, 1.e-05, 1.e-04, 1.e-03, 1.e-02, 1.e-01, 1.e+00, 1.e+01,
       1.e+02, 1.e+03, 1.e+04, 1.e+05, 1.e+06]),
 'vect__max_df': (0.2, 0.4, 0.6, 0.8, 1.0),
 'vect__min_df': (1, 3, 5, 10),
 'vect__ngram_range': ((1, 1), (1, 2)),
 'vect__norm': ('l1', 'l2')}
from time import time

t0 = time(),
print(f"Done in {time() - t0:.3f}s")
Fitting 5 folds for each of 40 candidates, totalling 200 fits
Done in 26.426s
print("Best parameters combination found:")
best_parameters = random_search.best_estimator_.get_params()
for param_name in sorted(parameter_grid.keys()):
    print(f"{param_name}: {best_parameters[param_name]}")
Best parameters combination found:
clf__alpha: 0.01
vect__max_df: 0.2
vect__min_df: 1
vect__ngram_range: (1, 1)
vect__norm: l1
test_accuracy = random_search.score(,
    "Accuracy of the best parameters using the inner CV of "
    f"the random search: {random_search.best_score_:.3f}"
print(f"Accuracy on test set: {test_accuracy:.3f}")
Accuracy of the best parameters using the inner CV of the random search: 0.816
Accuracy on test set: 0.709

The prefixes vect and clf are required to avoid possible ambiguities in the pipeline, but are not necessary for visualizing the results. Because of this, we define a function that will rename the tuned hyperparameters and improve the readability.

import pandas as pd

def shorten_param(param_name):
    """Remove components' prefixes in param_name."""
    if "__" in param_name:
        return param_name.rsplit("__", 1)[1]
    return param_name

cv_results = pd.DataFrame(random_search.cv_results_)
cv_results = cv_results.rename(shorten_param, axis=1)

We can use a to visualize the trade-off between scoring time and mean test score (i.e. “CV score”). Passing the cursor over a given point displays the corresponding parameters. Error bars correspond to one standard deviation as computed in the different folds of the cross-validation.

import as px

param_names = [shorten_param(name) for name in parameter_grid.keys()]
labels = {
    "mean_score_time": "CV Score time (s)",
    "mean_test_score": "CV score (accuracy)",
fig = px.scatter(
        "text": "trade-off between scoring time and mean test score",
        "y": 0.95,
        "x": 0.5,
        "xanchor": "center",
        "yanchor": "top",

Notice that the cluster of models in the upper-left corner of the plot have the best trade-off between accuracy and scoring time. In this case, using bigrams increases the required scoring time without improving considerably the accuracy of the pipeline.


For more information on how to customize an automated tuning to maximize score and minimize scoring time, see the example notebook Custom refit strategy of a grid search with cross-validation.

We can also use a to further visualize the mean test score as a function of the tuned hyperparameters. This helps finding interactions between more than two hyperparameters and provide intuition on their relevance for improving the performance of a pipeline.

We apply a math.log10 transformation on the alpha axis to spread the active range and improve the readability of the plot. A value \(x\) on said axis is to be understood as \(10^x\).

import math

column_results = param_names + ["mean_test_score", "mean_score_time"]

transform_funcs = dict.fromkeys(column_results, lambda x: x)
# Using a logarithmic scale for alpha
transform_funcs["alpha"] = math.log10
# L1 norms are mapped to index 1, and L2 norms to index 2
transform_funcs["norm"] = lambda x: 2 if x == "l2" else 1
# Unigrams are mapped to index 1 and bigrams to index 2
transform_funcs["ngram_range"] = lambda x: x[1]

fig = px.parallel_coordinates(
        "text": "Parallel coordinates plot of text classifier pipeline",
        "y": 0.99,
        "x": 0.5,
        "xanchor": "center",
        "yanchor": "top",

The parallel coordinates plot displays the values of the hyperparameters on different columns while the performance metric is color coded. It is possible to select a range of results by clicking and holding on any axis of the parallel coordinate plot. You can then slide (move) the range selection and cross two selections to see the intersections. You can undo a selection by clicking once again on the same axis.

In particular for this hyperparameter search, it is interesting to notice that the top performing models do not seem to depend on the regularization norm, but they do depend on a trade-off between max_df, min_df and the regularization strength alpha. The reason is that including noisy features (i.e. max_df close to \(1.0\) or min_df close to \(0\)) tend to overfit and therefore require a stronger regularization to compensate. Having less features require less regularization and less scoring time.

The best accuracy scores are obtained when alpha is between \(10^{-6}\) and \(10^0\), regardless of the hyperparameter norm.

Total running time of the script: (0 minutes 28.407 seconds)

Related examples

Classification of text documents using sparse features

Classification of text documents using sparse features

Comparing Random Forests and Histogram Gradient Boosting models

Comparing Random Forests and Histogram Gradient Boosting models

Balance model complexity and cross-validated score

Balance model complexity and cross-validated score

Column Transformer with Mixed Types

Column Transformer with Mixed Types

Clustering text documents using k-means

Clustering text documents using k-means

Gallery generated by Sphinx-Gallery