Implementation:Interpretml Interpret Construct Bins

Field	Value
Sources	InterpretML
Domains	Data_Preprocessing, Interpretability
Last Updated	2026-02-07 12:00 GMT

Overview

Construct_Bins is a concrete tool for discretizing features into bins provided by the InterpretML library.

Description

The construct_bins function takes cleaned feature data and target values, then computes bin definitions (cut points for continuous features, category mappings for nominal features) using the specified binning strategy. It also returns bin weights, feature bounds, histogram data, and optional differential privacy noise scaling. This function supports multiple binning strategies including quantile-based, uniform, and humanized quantile binning, and can operate under differential privacy constraints when epsilon and delta parameters are provided.

Usage

Import this function when you need to convert continuous and categorical features into discrete bins for EBM training. This is the primary entry point for the binning pipeline.

Code Reference

Source Location

Repository: interpretml/interpret
File: python/interpret-core/interpret/utils/_preprocessor.py
Lines: 602--676

Signature

def construct_bins(
    X,
    y,
    sample_weight,
    feature_names_given,
    feature_types_given,
    max_bins_leveled,
    binning="quantile",
    min_samples_bin=1,
    min_unique_continuous=0,
    seed=None,
    epsilon=None,
    delta=None,
    composition=None,
    privacy_bounds=None,
):

Import

from interpret.utils._preprocessor import construct_bins

I/O Contract

Inputs

Name	Type	Required	Description
`X`	array-like	Yes	Feature matrix
`y`	array-like	Yes	Target values
`sample_weight`	array-like	No	Per-sample weights
`feature_names_given`	list	Yes	Names for each feature column
`feature_types_given`	list	Yes	Types for each feature (e.g., "continuous", "nominal")
`max_bins_leveled`	list[int]	Yes	Maximum number of bins per interaction level
`binning`	str	No	Binning strategy: "quantile", "uniform", or "quantile_humanized" (default: "quantile")
`min_samples_bin`	int	No	Minimum samples per bin (default: 1)
`min_unique_continuous`	int	No	Minimum unique values for continuous treatment (default: 0)
`seed`	int	No	Random seed for reproducibility
`epsilon`	float	No	Differential privacy epsilon parameter
`delta`	float	No	Differential privacy delta parameter
`composition`	str	No	DP composition method
`privacy_bounds`	dict	No	Per-feature privacy bounds

Outputs

Returns a tuple of:

Index	Name	Type	Description
0	`feature_names_in`	list	Validated feature names
1	`feature_types_in`	list	Validated feature types
2	`bins`	list	Bin definitions per feature
3	`bin_weights`	list	Weight tensors per bin
4	`feature_bounds`	list	Min/max bounds per feature
5	`histogram_weights`	list	Histogram weight data
6	`missing_val_counts`	list	Count of missing values per feature
7	`unique_val_counts`	list	Count of unique values per feature
8	`noise_scale`	float/None	DP noise scale (None if DP not used)

Usage Examples

Basic Usage

from interpret.utils._preprocessor import construct_bins
import numpy as np

X = np.array([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0], [7.0, 8.0]])
y = np.array([0, 1, 0, 1])

result = construct_bins(
    X, y,
    sample_weight=None,
    feature_names_given=["feature_0", "feature_1"],
    feature_types_given=["continuous", "continuous"],
    max_bins_leveled=[256],
    binning="quantile",
)
feature_names_in, feature_types_in, bins, bin_weights, feature_bounds, histogram_weights, missing_val_counts, unique_val_counts, noise_scale = result

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment