HistGradientBoostingClassifier are the histogram bins equal in "width" or dynamically set depending on the feature distribution #30073

HeloiseS · 2024-10-15T07:20:37Z

HeloiseS
Oct 15, 2024

Hello sklearn community!

I was wondering how the histograms were created and if it was possible to visualise them at all. You see a lot of my features have funky distributions with long tails (see the picture), which means a basic histogram where all the bins have the same widths will loose all the resolution at the low end of the values.

I assume the algos in HGBDT are smart and handle this (otherwise my models really wouldn't be working as well as they do), but it was wondering if:

1. should i be processing my features to not have the long tails
1. Is there a way to plot the histograms that sklearn as made to discretise my features

I promise I have tried to search for this but I get drowned in a sea of articles that just state that xgboost and sklearn use histograms, they don't explain how or how to visualise. Also sorry if it's a dumb question because i misunderstood something!

Thanks for your amazing work to bring these tools to the public and non experts!

gitsult4n · 2026-08-28T02:38:10Z

gitsult4n
Aug 28, 2026

Short answer
The bins are not equal width
They are quantile based so each bin holds roughly the same number of samples
Narrow bins where your data is dense and wide bins in the long tail

The internal _BinMapper docstring states it directly

Bin thresholds are computed with the quantiles so that each bin contains
approximately the same number of samples.

Thresholds come from np.percentile on up to 200000 randomly subsampled rows
A feature with at most max_bins distinct values gets the midpoints between its distinct values instead

1 should you preprocess the long tails
Usually no
Quantile binning adapts to the distribution and a monotonic transform like log would give essentially the same bins

2 how to see the bins
A fitted model keeps them in a private bin mapper

import numpy as np
import matplotlib.pyplot as plt
from sklearn.ensemble import HistGradientBoostingClassifier

clf = HistGradientBoostingClassifier().fit(X, y)

edges = clf._bin_mapper.bin_thresholds_[0]  # thresholds for feature 0
bins = np.r_[X[:, 0].min(), edges, X[:, 0].max()]
plt.hist(X[:, 0], bins=bins)
plt.xscale("log")
plt.show()

Each array holds up to max_bins - 1 thresholds which are the upper bounds of the bins
The last bin index is always reserved for missing values

Caveat
_bin_mapper is private API so it can change without deprecation
Checked on scikit-learn 1.9.0 in sklearn/ensemble/_hist_gradient_boosting/binning.py

0 replies

Navigation Menu

Uh oh!

Uh oh!

HistGradientBoostingClassifier are the histogram bins equal in "width" or dynamically set depending on the feature distribution #30073

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{editor}}'s edit

{{editor}}'s edit

Uh oh!

Replies: 1 comment

Uh oh!

{{title}}

Uh oh!

Select a reply

Uh oh!

Uh oh!

Uh oh!

HistGradientBoostingClassifier are the histogram bins equal in "width" or dynamically set depending on the feature distribution #30073

Uh oh!

Uh oh!

HeloiseS Oct 15, 2024

I promise I have tried to search for this but I get drowned in a sea of articles that just state that xgboost and sklearn use histograms, they don't explain how or how to visualise. Also sorry if it's a dumb question because i misunderstood something!

Replies: 1 comment

Uh oh!

gitsult4n Aug 28, 2026

HeloiseS
Oct 15, 2024

gitsult4n
Aug 28, 2026