Encoding: from a dataframe to a numerical matrix for machine learning

This example shows how to transform a rich dataframe with columns of various types into a numerical matrix on which machine-learning algorithms can be applied. We study the case of predicting wages using the employee salaries dataset.

Easy learning on a dataframe

Let’s first retrieve the dataset, using one of the downloaders from the skrub.datasets module. As all the downloaders, fetch_employee_salaries() returns a dataset with a path field pointing to the dataframe file, which contains both the features and the target. We load the dataframe from the path using pandas. X is a dataframe which contains the features (aka design matrix, explanatory variables, independent variables). y is a column (pandas Series) which contains the target (aka dependent, response variable) that we want to learn to predict from X. In this case y is the annual salary, found in column “current_annual_salary”.

import pandas as pd

from skrub.datasets import fetch_employee_salaries

file_path = fetch_employee_salaries().path
employees = pd.read_csv(file_path)
X = employees.drop(columns="current_annual_salary")
y = employees["current_annual_salary"]

Most machine-learning algorithms work with arrays of numbers. The challenge here is that the employees dataframe is a heterogeneous set of columns: some are numerical ('year_first_hired'), some dates ('date_first_hired'), some have a few categorical entries ('gender'), some many ('employee_position_title'). Therefore our table needs to be “vectorized”: processed to extract numeric features.

skrub provides an easy way to build a simple but reliable machine-learning model which includes this step, working well on most tabular data.

from sklearn.model_selection import cross_validate

from skrub import tabular_pipeline

model = tabular_pipeline("regressor")
results = cross_validate(model, X, y)
results["test_score"]

array([0.90842776, 0.87932023, 0.91346596, 0.92178993, 0.9210775 ])

The estimator returned by tabular_pipeline combines 2 steps:

• a TableVectorizer to preprocess the dataframe and vectorize the features

• a supervised learner (by default a HistGradientBoostingRegressor)

model

Pipeline(steps=[('tablevectorizer',
                 TableVectorizer(low_cardinality=ToCategorical())),
                ('histgradientboostingregressor',
                 HistGradientBoostingRegressor())])

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 nbviewer.org.

In the rest of this example, we focus on the first step and explore the capabilities of skrub’s TableVectorizer.

More details on encoding tabular data

from skrub import TableVectorizer

vectorizer = TableVectorizer()
vectorized_X = vectorizer.fit_transform(X)
vectorized_X

Please enable javascript

The skrub table reports need javascript to display correctly. If you are displaying a report in a Jupyter notebook and you see this message, you may need to re-execute the cell or to trust the notebook (button on the top right or "File > Trust notebook").

From our 8 columns, the TableVectorizer has extracted 143 numerical features. Most of them are one-hot encoded representations of the categorical features. For example, we can see that 3 columns 'gender_F', 'gender_M', 'gender_nan' were created to encode the 'gender' column.

By performing appropriate transformations on our complex data, the TableVectorizer produced numeric features that we can use for machine-learning:

from sklearn.ensemble import HistGradientBoostingRegressor

HistGradientBoostingRegressor().fit(vectorized_X, y)

HistGradientBoostingRegressor()

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 nbviewer.org.

The TableVectorizer bridges the gap between tabular data and machine-learning pipelines. It allows us to apply a machine-learning estimator to our dataframe without manual data wrangling and feature extraction.

Inspecting the TableVectorizer

The TableVectorizer distinguishes between 4 basic kinds of columns (more may be added in the future). For each kind, it applies a different transformation, which we can configure. The kinds of columns and the default transformation for each of them are:

• numeric columns: simply casting to floating-point

• datetime columns: extracting features such as year, day, hour with the DatetimeEncoder

• low-cardinality categorical columns: one-hot encoding

• high-cardinality categorical columns: a simple and effective text representation pipeline provided by the GapEncoder

vectorizer

TableVectorizer()

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 nbviewer.org.

We can inspect which transformation was chosen for each column and retrieve the fitted transformer. vectorizer.kind_to_columns_ provides an overview of how the vectorizer categorized columns in our input:

vectorizer.kind_to_columns_

{'numeric': ['year_first_hired'], 'datetime': ['date_first_hired'], 'low_cardinality': ['gender', 'department', 'department_name', 'assignment_category'], 'high_cardinality': ['division', 'employee_position_title'], 'specific': []}

The reverse mapping is given by:

vectorizer.column_to_kind_

{'year_first_hired': 'numeric', 'date_first_hired': 'datetime', 'gender': 'low_cardinality', 'department': 'low_cardinality', 'department_name': 'low_cardinality', 'assignment_category': 'low_cardinality', 'division': 'high_cardinality', 'employee_position_title': 'high_cardinality'}

vectorizer.transformers_ gives us a dictionary which maps column names to the corresponding transformer.

vectorizer.transformers_["date_first_hired"]

DatetimeEncoder()

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 nbviewer.org.

We can also see which features in the vectorizer’s output were derived from a given input column.

vectorizer.input_to_outputs_["date_first_hired"]

['date_first_hired_year', 'date_first_hired_month', 'date_first_hired_day', 'date_first_hired_total_seconds']

vectorized_X[vectorizer.input_to_outputs_["date_first_hired"]]

Please enable javascript

The skrub table reports need javascript to display correctly. If you are displaying a report in a Jupyter notebook and you see this message, you may need to re-execute the cell or to trust the notebook (button on the top right or "File > Trust notebook").

Finally, we can go in the opposite direction: given a column in the input, find out from which input column it was derived.

vectorizer.output_to_input_["department_BOA"]

'department'

Dataframe preprocessing

Note that "date_first_hired" has been recognized and processed as a datetime column.

vectorizer.column_to_kind_["date_first_hired"]

'datetime'

But looking closer at our original dataframe, it was encoded as a string.

X["date_first_hired"]

0       09/22/1986
1       09/12/1988
2       11/19/1989
3       05/05/2014
4       03/05/2007
           ...
9223    11/03/2015
9224    11/28/1988
9225    04/30/2001
9226    09/05/2006
9227    01/30/2012
Name: date_first_hired, Length: 9228, dtype: str

Note the dtype: object in the output above. Before applying the transformers we specify, the TableVectorizer performs a few preprocessing steps.

For example, strings commonly used to represent missing values such as "N/A" are replaced with actual null. As we saw above, columns containing strings that represent dates (e.g. '2024-05-15') are detected and converted to proper datetimes.

We can inspect the list of steps that were applied to a given column:

vectorizer.all_processing_steps_["date_first_hired"]

[CleanNullStrings(), DropUninformative(), ToDatetime(), DatetimeEncoder(), {'date_first_hired_day': ToFloat(), 'date_first_hired_month': ToFloat(), ...}]

These preprocessing steps depend on the column:

vectorizer.all_processing_steps_["department"]

[CleanNullStrings(), DropUninformative(), ToStr(), OneHotEncoder(drop='if_binary', dtype='float32', handle_unknown='ignore',
              sparse_output=False), {'department_BOA': ToFloat(), 'department_BOE': ToFloat(), ...}]

A simple Pipeline for tabular data

The TableVectorizer outputs data that can be understood by a scikit-learn estimator. Therefore we can easily build a 2-step scikit-learn Pipeline that we can fit, test or cross-validate and that works well on tabular data.

import numpy as np
from sklearn.ensemble import HistGradientBoostingRegressor
from sklearn.model_selection import cross_validate
from sklearn.pipeline import make_pipeline

pipeline = make_pipeline(TableVectorizer(), HistGradientBoostingRegressor())

results = cross_validate(pipeline, X, y)
scores = results["test_score"]
print(f"R2 score:  mean: {np.mean(scores):.3f}; std: {np.std(scores):.3f}")
print(f"mean fit time: {np.mean(results['fit_time']):.3f} seconds")

R2 score:  mean: 0.912; std: 0.015
mean fit time: 1.628 seconds

Specializing the TableVectorizer for HistGradientBoosting

The encoders used by default by the TableVectorizer are safe choices for a wide range of downstream estimators. If we know we want to use it with a HistGradientBoostingRegressor (or classifier) model, we can make some different choices that are only well-suited for tree-based models but can yield a faster pipeline. We make 2 changes.

The HistGradientBoostingRegressor has built-in support for categorical features, so we do not need to one-hot encode them. We do need to tell it which features should be treated as categorical with the categorical_features parameter. In recent versions of scikit-learn, we can set categorical_features='from_dtype', and it will treat all columns in the input that have a Categorical dtype as such. Therefore we change the encoder for low-cardinality columns: instead of OneHotEncoder, we use skrub’s ToCategorical. This transformer will simply ensure our columns have an actual Categorical dtype (as opposed to string for example), so that they can be recognized by the HistGradientBoostingRegressor.

The second change replaces the GapEncoder with a MinHashEncoder. The GapEncoder is a topic model. It produces interpretable embeddings in a vector space where distances are meaningful, which is great for interpretation and necessary for some downstream supervised learners such as linear models. However fitting the topic model is costly in computation time and memory. The MinHashEncoder produces features that are not easy to interpret, but that decision trees can efficiently use to test for the occurrence of particular character n-grams (more details are provided in its documentation). Therefore it can be a faster and very effective alternative, when the supervised learner is built on top of decision trees, which is the case for the HistGradientBoostingRegressor.

The resulting pipeline is identical to the one produced by default by tabular_pipeline.

from skrub import MinHashEncoder, ToCategorical

vectorizer = TableVectorizer(
    low_cardinality=ToCategorical(), high_cardinality=MinHashEncoder()
)
pipeline = make_pipeline(
    vectorizer, HistGradientBoostingRegressor(categorical_features="from_dtype")
)

results = cross_validate(pipeline, X, y)
scores = results["test_score"]
print(f"R2 score:  mean: {np.mean(scores):.3f}; std: {np.std(scores):.3f}")
print(f"mean fit time: {np.mean(results['fit_time']):.3f} seconds")

R2 score:  mean: 0.916; std: 0.011
mean fit time: 0.950 seconds

We can see that this new pipeline achieves a similar score but is fitted much faster. This is mostly due to replacing GapEncoder with MinHashEncoder (however this makes the features less interpretable).

Feature importances in the statistical model

As we just saw, we can fit a MinHashEncoder faster than a GapEncoder. However, the GapEncoder has a crucial advantage: each dimension of its output space is associated with a topic which can be inspected and interpreted. In this section, after training a regressor, we will plot the feature importances.

Note:

To minimize computation time, we use the feature importances computed by the RandomForestRegressor, but you should prefer permutation_importance() instead (which are less subject to biases).

First, we train another scikit-learn regressor, the RandomForestRegressor:

from sklearn.ensemble import RandomForestRegressor

vectorizer = TableVectorizer()  # now using the default GapEncoder
regressor = RandomForestRegressor(n_estimators=50, max_depth=20, random_state=0)

pipeline = make_pipeline(vectorizer, regressor)
pipeline.fit(X, y)

Pipeline(steps=[('tablevectorizer', TableVectorizer()),
                ('randomforestregressor',
                 RandomForestRegressor(max_depth=20, n_estimators=50,
                                       random_state=0))])

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 nbviewer.org.

We are retrieving the feature importances:

avg_importances = regressor.feature_importances_
std_importances = np.std(
    [tree.feature_importances_ for tree in regressor.estimators_], axis=0
)
indices = np.argsort(avg_importances)[::-1]

And plotting the results:

import matplotlib.pyplot as plt

top_indices = indices[:20]
labels = vectorizer.get_feature_names_out()[top_indices]

plt.figure(figsize=(12, 9))
plt.barh(
    y=labels,
    width=avg_importances[top_indices],
    xerr=std_importances[top_indices],
    ecolor="k",
    color="b",
    alpha=0.5,
)
plt.yticks(fontsize=15)
plt.title("Feature importances")
plt.tight_layout(pad=1)
plt.show()

The GapEncoder creates feature names that show the first 3 most important words in the topic associated with each feature. As we can see in the plot above, this helps inspecting the model. If we had used a MinHashEncoder instead, the features would be much less helpful, with names such as employee_position_title_0, employee_position_title_1, etc.

We can see that features such the time elapsed since being hired, having a full-time employment, and the position, seem to be the most informative for prediction. However, feature importances must not be over-interpreted – they capture statistical associations rather than causal effects. Moreover, the fast feature importance method used here suffers from biases favouring features with larger cardinality, as illustrated in a scikit-learn example. In general we should prefer permutation_importance(), but it is a slower method.

Conclusion

In this example, we motivated the need for a simple machine learning pipeline, which we built using the TableVectorizer and a HistGradientBoostingRegressor.

We saw that by default, it works well on a heterogeneous dataset.

To better understand our dataset, and without much effort, we were also able to plot the feature importances.

Estimated memory usage: 611 MB