[FEAT] Add ALE (old fork) by abaziree · Pull Request #689 · mind-inria/hidimstat

abaziree · 2026-06-18T11:45:51Z

Issue #297

I added 1D ALE for continuous and discrete numeric features, and 2D ALE.
There are currently no tests.

Reference paper: https://arxiv.org/pdf/1612.08468

Use of AI tools:

LLM tools were used to generate at least part of the content of this pull request.

codecov · 2026-06-18T11:55:38Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.16%. Comparing base (60c4194) to head (c025899).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #689   +/-   ##
=======================================
  Coverage   98.16%   98.16%           
=======================================
  Files          23       23           
  Lines        1633     1633           
=======================================
  Hits         1603     1603           
  Misses         30       30

☔ View full report in Codecov by Harness.
📢 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.

bthirion

I'm still wondering what we should do for discrete features. Isn't PDP a more natural solution for such feature ? I'm worried about the arbitrary ordering, that may lead to a biased representation.

To be discussed again.

bthirion · 2026-06-18T19:42:04Z

+        to gather local samples in each bin.
+    feature_idx : int
+        Column index of the feature of interest.
+    grid_resolution : int, default=50


could have an auto value, where the number of bins is set in order to minimize the histogram error.

bthirion · 2026-06-18T19:44:02Z

+    difference in model output when the feature moves from the lower to the upper
+    neighbour values across all samples that have that value. The cumulative sum
+    of these average effects gives the (uncentered) ALE curve, which is then
+    centered so its weighted mean is zero.


Is there a reference to justify the behavior ?

The docstring is maybe not precise enough. In an interval ]x_{k-1}, x_k] of the feature of interest , we calculate the local effect by considering on the one hand samples with value x_{k-1} subsequently shifted to the value x_k, and on the other hand samples with value x_k subsequently shifted to the value x_{k-1}. It's exactly the same formula as the one in the Apley & Zhu paper.

Please include a ref to that paper.

bthirion · 2026-06-18T19:44:45Z

+    X : array-like of shape (n_samples, n_features)
+    feature_indices : tuple or list of two ints
+        Column indices ``[i, j]`` of the two features of interest.
+    grid_resolution : int, default=20


could also have an auto mode

bthirion · 2026-06-18T19:46:48Z

+                \frac{\partial f(X)}{\partial X_j}
+                \;\middle|\; X_j = z
+            \right] dz - c
+


We should include a bilbiographic reference to the definition

I added the Apley and Zhu reference. Should we also add Molnar's one even though I haven't specifically used it ?

GrituX · 2026-06-19T12:47:26Z

+def _check_estimator_predict(estimator):
+    """Return a scalar-output prediction callable for *estimator*.
+
+    For classifiers we use ``predict_proba`` (returns the probability of the
+    positive class for binary problems) when available, otherwise
+    ``decision_function``.  For regressors we use ``predict``.
+
+    Parameters
+    ----------
+    estimator : fitted sklearn-compatible estimator
+
+    Returns
+    -------
+    predict_fn : callable
+        A function that accepts an array of shape ``(n_samples, n_features)``
+        and returns a 1D array of shape ``(n_samples,)``.
+    """
+    feature_names = getattr(estimator, "feature_names_in_", None)
+
+    if hasattr(estimator, "predict_proba"):
+
+        def predict_fn(X):
+            if feature_names is not None:
+                X = pd.DataFrame(X, columns=feature_names)
+            proba = estimator.predict_proba(X)
+            # Binary: keep only the positive-class column
+            if proba.ndim == 2 and proba.shape[1] == 2:
+                return proba[:, 1]
+            return proba
+
+        return predict_fn
+
+    if hasattr(estimator, "decision_function"):
+
+        def predict_fn(X):
+            if feature_names is not None:
+                X = pd.DataFrame(X, columns=feature_names)
+            return estimator.decision_function(X).ravel()
+
+        return predict_fn
+
+    if hasattr(estimator, "predict"):
+
+        def predict_fn(X):
+            if feature_names is not None:
+                X = pd.DataFrame(X, columns=feature_names)
+            return estimator.predict(X).ravel()
+
+        return predict_fn
+
+    raise ValueError(
+        "'estimator' must expose at least one of predict_proba, "
+        "decision_function, or predict."
+    )


This function has a few issues :

L38-54 can be refactored with getattr(estimator, prediction_func)(X).ravel()

imho, nested functions reduce code readability. Here, the function could directly return the prediction for instance

Yep, I agree with your second point, it's not really useful here.

GrituX · 2026-06-19T12:59:44Z

+    quantiles_i = _build_quantile_grid(x_i, grid_resolution)
+    quantiles_j = _build_quantile_grid(x_j, grid_resolution)
+
+    n_bins_i = len(quantiles_i) - 1
+    n_bins_j = len(quantiles_j) - 1
+
+    if n_bins_i < 1:
+        raise ValueError(
+            f"Feature {idx_i} has fewer than 2 unique quantile edges. Increase grid_resolution or check your data."
+        )
+    if n_bins_j < 1:
+        raise ValueError(
+            f"Feature {idx_j} has fewer than 2 unique quantile edges. Increase grid_resolution or check your data."
+        )
+
+    bin_idx_i = _bin_indices(x_i, quantiles_i)
+    bin_idx_j = _bin_indices(x_j, quantiles_j)
+
+    # Second-order finite differences: evaluate at the four corners of each 2D bin
+    X_ll = X.copy()
+    X_ll[:, idx_i] = quantiles_i[bin_idx_i]
+    X_ll[:, idx_j] = quantiles_j[bin_idx_j]
+
+    X_lh = X.copy()
+    X_lh[:, idx_i] = quantiles_i[bin_idx_i]
+    X_lh[:, idx_j] = quantiles_j[bin_idx_j + 1]
+
+    X_hl = X.copy()
+    X_hl[:, idx_i] = quantiles_i[bin_idx_i + 1]
+    X_hl[:, idx_j] = quantiles_j[bin_idx_j]
+
+    X_hh = X.copy()
+    X_hh[:, idx_i] = quantiles_i[bin_idx_i + 1]
+    X_hh[:, idx_j] = quantiles_j[bin_idx_j + 1]
+
+    # Second-order local effects
+    local_effects = (predict_fn(X_hh) - predict_fn(X_hl)) - (
+        predict_fn(X_lh) - predict_fn(X_ll)
+    )  # shape (n_samples,)


bin_idx could be a 1D array which could refactor this code with a loop, and reduce the number of copies of X

abaziree · 2026-06-24T17:44:32Z

Kept as draft to have a history of discussions

bthirion reviewed Jun 18, 2026

View reviewed changes

GrituX reviewed Jun 19, 2026

View reviewed changes

abaziree force-pushed the 297-ale branch from c025899 to 0c59c84 Compare June 23, 2026 01:37

abaziree added 12 commits June 24, 2026 13:37

feat: add ALE implementation

2497d4b

[FEAT] Added Accumulated Local Effects for discrete feature

b7c7a74

Code formatting

0fc3c6a

Fix typo in ALE docstring

0ab6e20

fixes and add confidence intervals

65db0c1

add tests for ALE

dba1748

add examples for ALE

01740f3

add Apley2020 reference

bfa6046

fix import order

72b99e7

remove old file

24d97a3

fix if condition

3fde102

add tests for ALE

5c3b646

abaziree force-pushed the 297-ale branch from 648ba55 to 5c3b646 Compare June 24, 2026 11:40

fixes

27ae0f4

abaziree marked this pull request as draft June 24, 2026 17:43

abaziree changed the title ~~[FEAT] Add ALE~~ [FEAT] Add ALE (old fork) Jun 24, 2026

Uh oh!

Conversation

abaziree commented Jun 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov Bot commented Jun 18, 2026

Codecov Report

Uh oh!

bthirion left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

abaziree commented Jun 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

abaziree commented Jun 18, 2026 •

edited

Loading