edp-tool — Empirical Dependence Plot (EDP)

A small exploratory-data-analysis tool that shows how a target variable behaves across grouped values of a feature. It is designed with categorical / binary targets in mind (it plots the observed class rate), but it also works for continuous targets (it plots the mean).

EDP vs. PDP — why the name changed

This tool was originally called a Partial Dependence Plot (PDP), but that name is technically inaccurate. A true PDP requires a fitted model and marginalizes its predictions over the other features.

The EDP uses no model and does no marginalization — it plots the observed conditional mean of the target given a single feature, straight from the data. That makes it closely related to the M-plot (marginal plot) from the ALE literature (Apley & Zhu, 2020). Calling it Empirical Dependence Plot reflects exactly what it computes.

The old pdp function still works as a deprecated alias (see below).

Install

Just copy edp_tool.py next to your notebook, or from Colab:

import os
url = 'https://raw.githubusercontent.com/attilalr/edp-tool/main/edp_tool.py'
if not os.path.isfile('edp_tool.py'):
    !wget -q {url}
from edp_tool import edp

Requirements: numpy, pandas, matplotlib (see requirements.txt).

Usage

from edp_tool import edp

# Binary target -> positive-class rate per bin, with a Wilson CI band
edp(df, ['age', 'income'], 'converted')

# Multiclass target -> one line per class
edp(df, ['petal length (cm)'], 'species')

# Continuous target -> mean per bin, with a standard-error band
edp(df, ['age'], 'price')

# Save figures instead of showing them
edp(df, features, 'target', writefolder='figs')

edp() returns a list of dicts (feature, fig, ax, and path when saved), so you can post-process or embed the figures.

Key parameters

Multiclass targets are handled natively — no manual one-hot encoding needed.

What's new in this version

• Renamed to EDP (edp_tool.edp); pdp_tool.pdp kept as a deprecated alias.
• Fixed the dead categorical branch (feature type is now detected correctly).
• Fixed n leaking across features and the maximum value being dropped from the last bin.
• Native multiclass support, Wilson confidence intervals for rates, optional baseline line and uniform binning.
• Figures are returned and properly closed; validation raises real exceptions.

Development

pip install -r requirements.txt pytest
pytest

License

MIT — see LICENSE.