The goal of this exercise is to evaluate the impact of feature preprocessing on a pipeline that uses a decision-tree-based classifier instead of logistic regression.
The first question is to empirically evaluate whether scaling numerical feature is helpful or not;
The second question is to evaluate whether it is empirically better (both from a computational and a statistical perspective) to use integer coded or one-hot encoded categories.
Hint: HistGradientBoostingClassifier does not yet support sparse input data. You might want to use
OneHotEncoder(handle_unknown="ignore", sparse=False) to force the use a dense representation as a workaround.
import pandas as pd
from sklearn.model_selection import cross_val_score
from sklearn.pipeline import make_pipeline
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import OrdinalEncoder
from sklearn.experimental import enable_hist_gradient_boosting
from sklearn.ensemble import HistGradientBoostingClassifier
df = pd.read_csv("https://www.openml.org/data/get_csv/1595261/adult-census.csv")
# Or use the local copy:
# df = pd.read_csv('../datasets/adult-census.csv')
target_name = "class"
target = df[target_name].to_numpy()
data = df.drop(columns=[target_name, "fnlwgt"])
numerical_columns = [c for c in data.columns
if data[c].dtype.kind in ["i", "f"]]
categorical_columns = [c for c in data.columns
if data[c].dtype.kind not in ["i", "f"]]
categories = [data[column].unique()
for column in data[categorical_columns]]
First let's time the pipeline we used in the main notebook to serve as a reference:
%%time
preprocessor = ColumnTransformer([
('categorical', OrdinalEncoder(categories=categories), categorical_columns),
], remainder="passthrough")
model = make_pipeline(
preprocessor,
HistGradientBoostingClassifier()
)
scores = cross_val_score(model, data, target)
print(f"The different scores obtained are: \n{scores}")
print(f"The accuracy is: {scores.mean():.3f} +- {scores.std():.3f}")
The different scores obtained are: [0.87224895 0.87143003 0.8745905 0.87346437 0.87796888] The accuracy is: 0.874 +- 0.002 CPU times: user 12.8 s, sys: 138 ms, total: 12.9 s Wall time: 3.72 s
%%time
from sklearn.preprocessing import StandardScaler
preprocessor = ColumnTransformer([
('numerical', StandardScaler(), numerical_columns),
('categorical', OrdinalEncoder(categories=categories), categorical_columns),
])
model = make_pipeline(
preprocessor,
HistGradientBoostingClassifier()
)
scores = cross_val_score(model, data, target)
print(f"The different scores obtained are: \n{scores}")
print(f"The accuracy is: {scores.mean():.3f} +- {scores.std():.3f}")
The different scores obtained are: [0.87224895 0.87143003 0.8745905 0.87346437 0.87796888] The accuracy is: 0.874 +- 0.002 CPU times: user 15.8 s, sys: 103 ms, total: 15.9 s Wall time: 4.95 s
We can observe that both the accuracy and the training time are approximately the same as the reference pipeline (any time difference you might observe is not significant).
Scaling numerical features is indeed useless for most decision tree models in general and for HistGradientBoostingClassifier in particular.
For linear models, we have observed that integer coding of categorical
variables can be very detrimental. However for
HistGradientBoostingClassifier models, it does not seem to be the
case as the cross-validation of the reference pipeline with
OrdinalEncoder is good.
Let's see if we can get an even better accuracy with OneHotEncoding:
%%time
from sklearn.preprocessing import OneHotEncoder
preprocessor = ColumnTransformer([
('categorical',
OneHotEncoder(handle_unknown="ignore", sparse=False),
categorical_columns),
], remainder="passthrough")
model = make_pipeline(
preprocessor,
HistGradientBoostingClassifier()
)
scores = cross_val_score(model, data, target)
print(f"The different scores obtained are: \n{scores}")
print(f"The accuracy is: {scores.mean():.3f} +- {scores.std():.3f}")
The different scores obtained are: [0.87286314 0.87173713 0.87407862 0.87407862 0.87807125] The accuracy is: 0.874 +- 0.002 CPU times: user 42.7 s, sys: 830 ms, total: 43.5 s Wall time: 13.3 s
From an accuracy point of view, the result is almost exactly the same.
The reason is that HistGradientBoostingClassifier is expressive
and robust enough to deal with misleading ordering of integer coded
categories (which was not the case for linear models).
However from a computation point of view, the training time is
significantly longer: this is caused by the fact that OneHotEncoder
generates approximately 10 times more features than OrdinalEncoder.
Note that the current implementation HistGradientBoostingClassifier
is still incomplete, and once sparse representation are handled
correctly, training time might improve with such kinds of encodings.
The main take away message is that arbitrary integer coding of
categories is perfectly fine for HistGradientBoostingClassifier
and yields fast training times.