data: fix Marginal.visualize crash on categorical features (#119) by jbbqqf · Pull Request #666 · interpretml/interpret

jbbqqf · 2026-05-09T18:23:43Z

Summary

Marginal(...).explain_data(...).visualize(key) crashed with
TypeError: unsupported operand type(s) for -: 'str' and 'str' whenever
feature_types[key] was "nominal" or "ordinal", making the marginal
data explainer unusable on every dataset that contained a categorical
feature.

Fixes #119 — visualize has a bug for categorical features

Context

MarginalExplanation.visualize builds a small two-panel plotly figure
(feature histogram + response histogram). It computed the histogram
bin width for both panels with a single line at the top of the function:

bin_size = density_dict["names"][1] - density_dict["names"][0]

For continuous features density_dict["names"] is the numeric output
of np.histogram(...) (bin edges), so the subtraction is fine. For
nominal/ordinal features the same field stores string category labels
returned by np.unique(...), and the subtraction blows up before any
branch on is_categorical ever executes — so neither path renders.

There is a second instance of the same pattern further down for the
response axis (resp_bin_size = density_dict["names"][...] — note that
it even pointed at the feature density rather than the response
density). Both are addressed in this PR.

The fix matches the long-standing report from issue #119, with one
small addition: classification targets (categorical response) now also
render, because the response side is now guarded the same way.

Changes

python/interpret-core/interpret/data/_response.py
- move bin_size into the else branch so it is only computed when
  density_dict["names"] are numeric bin edges.
- mirror the same guard on the response axis. When the response is
  categorical (e.g. classification target) build a plotly Histogram
  without ybins, otherwise compute resp_bin_size from
  resp_density_dict["names"] (was previously computed off the
  feature density — a latent bug the same line fixed).
- add a tiny _is_numeric_bin_edges predicate so the categorical
  response check stays self-documenting; comment cites issue visualize has a bug for categorical features #119
  so future readers don't have to re-derive the reasoning.
python/interpret-core/tests/data/test_marginal_visualize.py (new)
- four pytest cases: nominal feature on continuous target, continuous
  feature (non-regression), nominal feature on classification target,
  and the key=None overall path.
- the two nominal-feature cases fail on origin/main with the exact
  TypeError cited in visualize has a bug for categorical features #119; all four pass on this branch.

Reproduce BEFORE/AFTER yourself (copy-paste)

# --- one-time setup ---
git clone https://github.com/interpretml/interpret.git /tmp/repro && cd /tmp/repro
python -m venv .venv && source .venv/bin/activate
pip install -e "python/interpret-core[testing]" plotly

cat > /tmp/repro_119.py <<'PY'
import numpy as np
import pandas as pd
from interpret.data import Marginal

df = pd.DataFrame({
    "cat":  ["a", "b", "a", "c", "b", "a", "c", "b"],
    "cont": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0],
})
y = np.array([0.5, 1.5, 2.5, 3.5, 4.5, 5.5, 6.5, 7.5])
m = Marginal(feature_types=["nominal", "continuous"]).explain_data(df, y)
print("categorical viz:", type(m.visualize(0)).__name__)
print("continuous viz: ", type(m.visualize(1)).__name__)
PY

# --- BEFORE (origin/main) ---
git checkout origin/main
python /tmp/repro_119.py
# Expected: TypeError: unsupported operand type(s) for -: 'str' and 'str'

# --- AFTER (this PR) ---
git fetch https://github.com/jbbqqf/interpret.git feat/119-fix-marginal-categorical-visualize
git checkout FETCH_HEAD
python /tmp/repro_119.py
# Expected: both lines print "Figure"

# Optional: run the new regression suite, which fails on main and
# passes on this branch
git checkout origin/main
python -m pytest -vv python/interpret-core/tests/data/test_marginal_visualize.py
# Expected: 2 failed, 2 passed (the categorical cases hit the bug)
git checkout FETCH_HEAD
python -m pytest -vv python/interpret-core/tests/data/test_marginal_visualize.py
# Expected: 4 passed

What I ran locally

python -m pytest -vv tests/data/test_marginal_visualize.py → 4 passed
on this branch.
The same suite on origin/main → 2 failed, 2 passed (the two
failures are the exact TypeError cited in visualize has a bug for categorical features #119; the passing ones
cover the continuous and overall paths that were already correct).
ruff check interpret-core/interpret/data/_response.py → 6 errors
reported, identical to origin/main (all pre-existing
PLC0415/NPY002 flags in this file). My change introduces zero
new lint findings.
ruff format --check on the touched files → already formatted.

The selenium / dashboard / full visual suites need extra optional
dependencies (requests, dash, etc.) that aren't part of the
testing extra; they were not run.

Edge cases tested

#	Scenario	Input	Expected	Verified by
1	Nominal feature, continuous target	`feature_types=["nominal","continuous"]`, key=0	plotly Figure (was TypeError)	`test_visualize_nominal_feature_continuous_target_succeeds`
2	Continuous feature (regression guard)	same dataset, key=1	plotly Figure (still works)	`test_visualize_continuous_feature_still_works`
3	Nominal feature, classification target	y = `[0,1,0,1,0,1,0,1]`, key=0	plotly Figure (categorical response branch)	`test_visualize_nominal_feature_classification_target`
4	Overall view (`key=None`)	any	plot_density figure (path unchanged)	`test_visualize_overall_unaffected`

Risk / blast radius

The change is local to MarginalExplanation.visualize. It only adds
code paths that were previously unreachable (categorical features /
classification responses). Continuous-feature behaviour is byte-equal
to before — same plotly trace, same xbins dict. The new
_is_numeric_bin_edges helper is module-private. No public API
changes.

Release note

Fixed Marginal data explainer crashing with TypeError when visualising
nominal/ordinal features or classification targets (issue #119).

PR drafted with assistance from Claude Code. The change was reviewed
manually against interpretml/interpret's source and the cited issue.
The reproducer block above was used during development; it is the same
one a reviewer can paste verbatim.

…ml#119) Marginal(...).explain_data(...).visualize(key) raised TypeError: unsupported operand type(s) for -: 'str' and 'str' whenever feature_types[key] was nominal/ordinal because the histogram path unconditionally computed density_dict["names"][1] - density_dict["names"][0] for the bin size, even though for categorical features the names list holds string category labels rather than numeric bin edges. Compute bin_size lazily only on the numeric branch, and apply the same guard on the response side so that a categorical response (e.g. classification target) also renders without crashing. A small helper _is_numeric_bin_edges keeps the predicate explicit. Add tests/data/test_marginal_visualize.py with four cases covering the two regression branches plus the previously-working continuous and overall paths to lock in the contract. Co-Authored-By: Claude Code <noreply@anthropic.com> Signed-off-by: jbbqqf <jbaptiste.braun@gmail.com>

codecov · 2026-05-09T18:45:28Z

Codecov Report

❌ Patch coverage is 87.50000% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 67.18%. Comparing base (6c79a67) to head (e78b256).

Files with missing lines	Patch %	Lines
python/interpret-core/interpret/data/_response.py	87.50%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #666      +/-   ##
==========================================
+ Coverage   66.86%   67.18%   +0.32%     
==========================================
  Files          77       77              
  Lines       11724    11729       +5     
==========================================
+ Hits         7839     7880      +41     
+ Misses       3885     3849      -36

Flag	Coverage Δ
bdist_linux_311_python	`66.91% <87.50%> (+0.32%)`	⬆️
bdist_linux_312_python	`66.93% <87.50%> (+0.33%)`	⬆️
bdist_linux_313_python	`66.93% <87.50%> (+0.33%)`	⬆️
bdist_linux_314_python	`66.82% <87.50%> (+0.35%)`	⬆️
bdist_linuxarm_311_python	`66.95% <87.50%> (+0.33%)`	⬆️
bdist_linuxarm_312_python	`66.95% <87.50%> (+0.32%)`	⬆️
bdist_linuxarm_313_python	`66.93% <87.50%> (+0.33%)`	⬆️
bdist_linuxarm_314_python	`66.86% <87.50%> (+0.33%)`	⬆️
bdist_mac_311_python	`67.08% <87.50%> (+0.32%)`	⬆️
bdist_mac_312_python	`67.09% <87.50%> (+0.33%)`	⬆️
bdist_mac_313_python	`67.09% <87.50%> (+0.33%)`	⬆️
bdist_mac_314_python	`66.98% <87.50%> (+0.32%)`	⬆️
bdist_win_311_python	`67.09% <87.50%> (+0.32%)`	⬆️
bdist_win_312_python	`67.09% <87.50%> (+0.32%)`	⬆️
bdist_win_313_python	`67.09% <87.50%> (+0.32%)`	⬆️
bdist_win_314_python	`66.98% <87.50%> (+0.30%)`	⬆️
sdist_linux_311_python	`66.86% <87.50%> (+0.32%)`	⬆️
sdist_linux_312_python	`66.88% <87.50%> (+0.33%)`	⬆️
sdist_linux_313_python	`66.86% <87.50%> (+0.32%)`	⬆️
sdist_linux_314_python	`66.79% <87.50%> (+0.33%)`	⬆️
sdist_linuxarm_311_python	`66.87% <87.50%> (+0.32%)`	⬆️
sdist_linuxarm_312_python	`66.87% <87.50%> (+0.32%)`	⬆️
sdist_linuxarm_313_python	`66.87% <87.50%> (+0.32%)`	⬆️
sdist_linuxarm_314_python	`66.78% <87.50%> (+0.32%)`	⬆️
sdist_mac_311_python	`66.99% <87.50%> (+0.32%)`	⬆️
sdist_mac_312_python	`66.99% <87.50%> (+0.32%)`	⬆️
sdist_mac_313_python	`67.01% <87.50%> (+0.33%)`	⬆️
sdist_mac_314_python	`66.90% <87.50%> (+0.32%)`	⬆️
sdist_win_311_python	`67.11% <87.50%> (+0.33%)`	⬆️
sdist_win_312_python	`67.11% <87.50%> (+0.33%)`	⬆️
sdist_win_313_python	`67.11% <87.50%> (+0.33%)`	⬆️
sdist_win_314_python	`67.02% <87.50%> (+0.33%)`	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

paulbkoch · 2026-05-12T04:18:26Z

Very nice PR. Thanks @jbbqqf!

paulbkoch merged commit 7a43690 into interpretml:main May 12, 2026
67 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

data: fix Marginal.visualize crash on categorical features (#119)#666

data: fix Marginal.visualize crash on categorical features (#119)#666
paulbkoch merged 1 commit into
interpretml:mainfrom
jbbqqf:feat/119-fix-marginal-categorical-visualize

jbbqqf commented May 9, 2026

Uh oh!

codecov Bot commented May 9, 2026 •

edited

Loading

Uh oh!

Uh oh!

paulbkoch commented May 12, 2026

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

jbbqqf commented May 9, 2026

Summary

Context

Changes

Reproduce BEFORE/AFTER yourself (copy-paste)

What I ran locally

Edge cases tested

Risk / blast radius

Release note

Uh oh!

codecov Bot commented May 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Uh oh!

paulbkoch commented May 12, 2026

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

2 participants

codecov Bot commented May 9, 2026 •

edited

Loading