FEA add CumulativeAccuracyDisplay #28752

JosephBARBIERDARNAL · 2024-04-02T16:55:37Z

Reference Issue

What does this implement/fix?

creation of a CumulativeAccuracyDisplay class for plots

"The CAP of a model represents the cumulative number of positive outcomes along the y-axis versus the corresponding cumulative number of a classifying parameter along the x-axis. The output is called a CAP curve.[1] The CAP is distinct from the receiver operating characteristic (ROC) curve, which plots the true-positive rate against the false-positive rate." (wikipedia definition)

It's mainly inspired from the RocCurveDisplay class.

other

It's currently a work in progress.

github-actions · 2024-04-02T16:57:01Z

✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: 73ca739. Link to the linter CI: here}

sklearn/metrics/_plot/cap_curve.py

glemaitre

Thanks for the PR.

On a first pass, could you fix the linter CI. I would advise to directly use pre-commit. It would corresponds to the optional step 10 of the guideline(https://scikit-learn.org/dev/developers/contributing.html#how-to-contribute). It is optional but it makes life easier :)

ogrisel

Thanks very much for the PR. Here is quick first pass of reviews.

sklearn/metrics/_plot/cap_curve.py

ogrisel · 2024-04-02T17:32:16Z

sklearn/metrics/_plot/cap_curve.py

+
+CumulativeAccuracyDisplay.from_predictions(y_test, y_scores, name='Logistic Regression')
+plt.title('CAP Curve using from_predictions')
+plt.show()


For information when running this code I get:

This looks fine but I would like to also have an option to display the curve for oracle predictions and the curve for the random predictions baseline.

EDIT: we need to check the literature to see if it's common to normalize the cumulated values to 1 instead of plotting absolute numbers. Maybe an option to select between relative and absolute scales would make sense.

In addition, I think we should set better the x- and y-lim (as in ROC curve display).
Also, is it common for this curve to be be in a squared axis? I see this is kind of different from the ROC and PR curve in this regards, where the x- and y-axis are not the same unit.

An option for relative VS absolute scales sounds relevant to me. Don't have a opinion of which should be the default.

+1 for squared axis (especially with normalized scale). This paper (Tasche 2006 "Validation of internal rating systems and PD estimates" contains one example with a squared axis cap curve:

It reminds of ROC curve, but maybe there are reasons of not doing so?

An option for relative VS absolute scales sounds relevant to me. Don't have a opinion of which should be the default.

I would use relative intuitively as the default.

JosephBARBIERDARNAL · 2024-04-03T20:41:40Z

I've implemented some of the elements mentioned, but I don't have the time to do everything quickly.

I also have several concerns:

I haven't yet understood how to use this type of curve in the case of a regressor, and I don't think I've seen any article on the subject?
I'm not sure I understand how to use pre-commit/linters properly (I haven't managed to commit with pre-commit).
Being a total beginner in the use of tests, I haven't shifted the code for this, and could greatly use a simple practical explanation.

ogrisel · 2024-04-04T13:49:34Z

I haven't yet understood how to use this type of curve in the case of a regressor, and I don't think I've seen any article on the subject?

The example I linked above is precisely an example of a Lorenz curve which based on the linked discussion is very related to the CAP curve.

EDIT: I also posted an extra reference in the parent issue: #10003 (comment).

ogrisel · 2024-04-04T13:59:45Z

Being a total beginner in the use of tests, I haven't shifted the code for this, and could greatly use a simple practical explanation.

For an example, you can have a look at the tests for RocCurveDisplay in sklearn/metrics/_plot/tests/test_roc_curve_display.py.

You can run all the tests in this module with the pytest command (pip install pytest first if needed):

$ pytest sklearn/metrics/_plot/tests/test_roc_curve_display.py -vl

you can run a specific test with:

$ pytest sklearn/metrics/_plot/tests/test_roc_curve_display.py -vl -k test_roc_curve_display_plotting
================================================================================================ test session starts =================================================================================================
platform darwin -- Python 3.11.7, pytest-8.1.1, pluggy-1.4.0 -- /Users/ogrisel/miniforge3/envs/dev/bin/python3.11
cachedir: .pytest_cache
Using --randomly-seed=2393278150
hypothesis profile 'default' -> database=DirectoryBasedExampleDatabase(PosixPath('/Users/ogrisel/code/scikit-learn/.hypothesis/examples'))
benchmark: 4.0.0 (defaults: timer=time.perf_counter disable_gc=False min_rounds=5 min_time=0.000005 max_time=1.0 calibration_precision=10 warmup=False warmup_iterations=100000)
rootdir: /Users/ogrisel/code/scikit-learn
configfile: setup.cfg
plugins: snapshot-0.9.0, httpserver-1.0.8, repeat-0.9.3, randomly-3.15.0, cov-4.1.0, clarity-1.0.1, mock-3.12.0, hypothesis-6.93.0, xdist-3.5.0, anyio-4.2.0, benchmark-4.0.0
collected 53 items / 21 deselected / 32 selected                                                                                                                                                                     

sklearn/metrics/_plot/tests/test_roc_curve_display.py::test_roc_curve_display_plotting[from_estimator-LogisticRegression-False-False-True-decision_function] PASSED                                            [  3%]
sklearn/metrics/_plot/tests/test_roc_curve_display.py::test_roc_curve_display_plotting[from_predictions-Classifier-True-False-True-decision_function] PASSED                                                   [  6%]
sklearn/metrics/_plot/tests/test_roc_curve_display.py::test_roc_curve_display_plotting[from_estimator-LogisticRegression-False-True-True-predict_proba] PASSED                                                 [  9%]
sklearn/metrics/_plot/tests/test_roc_curve_display.py::test_roc_curve_display_plotting[from_estimator-LogisticRegression-False-False-False-decision_function] PASSED                                           [ 12%]
...
sklearn/metrics/_plot/tests/test_roc_curve_display.py::test_roc_curve_display_plotting[from_estimator-LogisticRegression-True-True-False-predict_proba] PASSED                                                 [ 96%]
sklearn/metrics/_plot/tests/test_roc_curve_display.py::test_roc_curve_display_plotting[from_estimator-LogisticRegression-True-True-True-predict_proba] PASSED                                                  [100%]

========================================================================================= 32 passed, 21 deselected in 0.67s ==========================================================================================

since this is a parametrized test (with @pytest.mark.parametrize decorators) this is runs actually 21 variations of the same test functions. To run a single version you can do:

$ pytest sklearn/metrics/_plot/tests/test_roc_curve_display.py -vl -k "test_roc_curve_display_plotting and from_estimator-LogisticRegression-True-True-True-predict_proba"

So for your branch, you can create a file named sklearn/metrics/_plot/tests/test_cap_curve_display.py or similar and write some test functions that make assertions similar to what we do for the ROC curve display.

More details on pytest usage in https://scikit-learn.org/dev/developers/tips.html#pytest-tips or on the pytest website itself.

ogrisel · 2024-04-04T14:14:53Z

I'm not sure I understand how to use pre-commit/linters properly (I haven't managed to commit with pre-commit).

Some pre-commit tasks such as the black formatted and the ruff import statement formatter can auto-fix the code with you type git commit. If that happens the commit command fails but the files have been fixed: you need to git add the auto-fixed files and then git commit again.

ogrisel

A few more comments:

sklearn/metrics/_plot/cap_curve.py

ogrisel · 2024-04-04T16:10:39Z

sklearn/metrics/_plot/cap_curve.py

+
+        # compute cumulative sums for true positives and all cases
+        cumulative_true_positives = np.cumsum(y_true_sorted * sample_weight_sorted)
+        cumulative_total = np.cumsum(sample_weight_sorted)


This is very similar to a Lorenz curve that we manually define in this example.

scikit-learn/examples/linear_model/plot_poisson_regression_non_normal_loss.py

Lines 495 to 507 in 6bf0ba5

def lorenz_curve(y_true, y_pred, exposure):

y_true, y_pred = np.asarray(y_true), np.asarray(y_pred)

exposure = np.asarray(exposure)

# order samples by increasing predicted risk:

ranking = np.argsort(y_pred)

ranked_frequencies = y_true[ranking]

ranked_exposure = exposure[ranking]

cumulated_claims = np.cumsum(ranked_frequencies * ranked_exposure)

cumulated_claims /= cumulated_claims[-1]

cumulated_exposure = np.cumsum(ranked_exposure)

cumulated_exposure /= cumulated_exposure[-1]

return cumulated_exposure, cumulated_claims

The differences are (as far as I can see):

y_pred is the estimated probability of a positive class membership given by a classifier for CAP/Lift/Gain curves while y_pred hold the continuous prediction of a positive target variable for a regressor.

sorted_indices = np.argsort(y_pred)[::-1] is reversed in CAP while it stays in ascending order in a Lorenz curve.

So I think we could indeed reuse the same class both for CAP/Lift/Gain curves and Lorenz curves.

To switch between both, we can probably introduce a keyword parameter x_ordering_style="cap" vs x_ordering_style="lorenz" for instance.

As discussed in the main thread of the PR, let's keep that for a follow-up PR.

JosephBARBIERDARNAL · 2024-04-04T16:55:12Z

I'm not sure I understand how to use pre-commit/linters properly (I haven't managed to commit with pre-commit).

Some pre-commit tasks such as the black formatted and the ruff import statement formatter can auto-fix the code with you type git commit. If that happens the commit command fails but the files have been fixed: you need to git add the auto-fixed files and then git commit again.

thanks! helps a lot actually (also for tests)

ogrisel · 2024-04-05T09:59:05Z

Great. Now that you have tests, you can see in the diff view that the test coverage analysis has resulted in a few automated inline comments that highlight section of your code that are currently not tested:

https://github.com/scikit-learn/scikit-learn/pull/28752/files

To merge this PR we will need to have improved test coverage.

However, I would rather put the focus on the upgrading the code itself to:

make it possible to use the display on positively valued regression tasks;
make it possible to swap the order of the x-axis to use a Lorenz curve style instead of a CAP/Lift curve style;
updated the linked example to use the display instead of reimplementing a Lorenz curve from scratch.

lorentzenchr · 2024-04-05T11:54:40Z

make it possible to swap the order of the x-axis to use a Lorenz curve style instead of a CAP/Lift curve style;

To keep this PR minimal, I would postpone this to a follow-up PR.

ogrisel · 2024-04-05T12:22:56Z

Alright, I agree to limit the scope of this PR to the case of CAP / Lift / Gain curve on classification tasks for now.

ogrisel

As I said above I am ok to focus this PR on the classification case and leave regression for later. However I think we should already chose public attribute names for the cumulated statistics that are compatible with the anticipated regression/Lorenz curve case (see inline comment below)

Furthermore, I think we should update one of the existing example that displays a ROC curve for a binary classifier to also include a CAP curve. I think that the following example that already displays both a DET curve and a ROC curve is a good candidate:

https://scikit-learn.org/stable/auto_examples/model_selection/plot_det.html

We should probably refactor it a bit to change the title and even the filename to make it explicit that it's about more than DET curves.

The alternative would be to create a new example file dedicated to CAP but I think we all agree that the example gallery has already too many short examples and it's better to consolidate them into larger examples that combine related components from the scikit-learn API and better contrasts the commonalities and differences.

ogrisel · 2024-04-05T12:30:34Z

sklearn/metrics/_plot/tests/test_cap_curve_display.py

+    if plot_chance_level:
+        assert (
+            cap_display.chance_level_ is not None
+        ), "Chance level line should be present"


Let's also extend the tests to check for the presence shape and dtypes of other public attributes such as y_true_cumulative (new name suggested in https://github.com/scikit-learn/scikit-learn/pull/28752/files#r1553497449).

I might also make sense to check that the values of those attribute match (with assert_allclose) when calling CumulativeAccuracyDisplay.from_estimator or CumulativeAccuracyDisplay.from_predictions.

I've made the change for y_true_cumulative, but I'm not sure about shape or dtypes. Are displays supposed to have these attributes (or just this one)?

I've just realised that I've misread it. You meant checking the attributes of y_true_cumulative, not the display itself.

sklearn/metrics/_plot/cap_curve.py

lorentzenchr

Review of from_estimator.

sklearn/metrics/_plot/cap_curve.py

lorentzenchr · 2024-04-05T13:33:15Z

sklearn/metrics/_plot/cap_curve.py

+        This is also known as a Gain or Lift Curve for classification, and a Lorenz
+        curve for regression with a positively valued target.


Suggested change

This is also known as a Gain or Lift Curve for classification, and a Lorenz

curve for regression with a positively valued target.

This is also known as a Gain or Lift Curve, or the mirrored Lorenz

curve of Machine Learning for regression with a positively valued target.

Since we agree to leave the Lorenz curve / regression case for a follow-up PR I would defer this change for later. Speaking about regression on an class that is designed and document to only accept classifier predictions would be a source confusion.

lorentzenchr · 2024-04-05T15:00:50Z

The alternative would be to create a new example file dedicated to CAP but I think we all agree that the example gallery has already too many short examples

I would also not add another example. We can replace the lorenz curve with CAP in 2 examples (Poisson and Tweedie).

ogrisel · 2024-04-08T08:49:28Z

We can replace the lorenz curve with CAP in 2 examples (Poisson and Tweedie).

But this is only for regression. I would like to illustrate CAP in a classification context and updating an example with a ROC and DET curve would make that possible and also improve the discoverability of the plot. Furthermore I find it interesting to visually compare the 3 plots for the same models / data.

ogrisel

The sklearn.metrics should expose the CumulativeAccuracyDisplay class as part of the public API and include it in its __all__ module level attribute.

I think the docstrings of the 2 class methods should:

mention that this curve assesses the ranking power of classifiers, that is the ability to correctly order the individual predictions by probability of actually be labeled by the positive class,
make it explicit this curve is invariant under any monotonic transformation of the predictions, therefore it cannot be used to assess the calibration of the predictions, only the ranking power,
have a "See also" section (at the end of the docstring) and cross-link RocCurveDisplay as an alternative curve to assess the ranking power of classifiers.

sklearn/metrics/_plot/cap_curve.py

lorentzenchr · 2024-04-08T09:25:24Z

I just want to avoid a new example. Therefore, either extend existing examples or postpone to a followup PR. But please, no new example! (Strong opinion there)

sklearn/metrics/_plot/cap_curve.py

lorentzenchr · 2024-04-08T10:27:43Z

Alright, I agree to limit the scope of this PR to the case of CAP / Lift / Gain curve on classification tasks for now.

I'm more in favor of including regression from the start. That's where it adds additional value. For classification, it is kind of equivalent to ROC. On top, this helps to find good names.

lorentzenchr · 2024-04-28T12:57:12Z

@JosephBARBIERDARNAL Why did you close? This would be a valuable addition.

JosephBARBIERDARNAL · 2024-04-28T15:33:07Z

I forgot that this PR was still open when I deleted my fork, sorry! Any solution for this type of case?

JosephBARBIERDARNAL · 2024-05-07T19:45:49Z

I created a new PR (#28972) with the same code. I'll hopefully keep this one open. again sorry for the mistake!

create cumulative accuracy profile class

2630953

github-actions bot added the module:metrics label Apr 2, 2024

JosephBARBIERDARNAL marked this pull request as draft April 2, 2024 16:56

glemaitre changed the title ~~create cumulative accuracy profile class~~ FEA add CumulativeAccuracyDisplay Apr 2, 2024

glemaitre reviewed Apr 2, 2024

View reviewed changes

sklearn/metrics/_plot/cap_curve.py Outdated Show resolved Hide resolved

glemaitre reviewed Apr 2, 2024

View reviewed changes

sklearn/metrics/_plot/cap_curve.py Outdated Show resolved Hide resolved

glemaitre reviewed Apr 2, 2024

View reviewed changes

ogrisel reviewed Apr 2, 2024

View reviewed changes

JosephBARBIERDARNAL and others added 2 commits April 3, 2024 22:31

CAP curve: add options for baseline and axis scale

5b42e01

Merge branch 'main' into cap-curve

cc674b0

Merge branch 'main' into cap-curve

aec617f

ogrisel reviewed Apr 4, 2024

View reviewed changes

create test file for cap curve display

73ca739

ogrisel reviewed Apr 5, 2024

View reviewed changes

sklearn/metrics/_plot/cap_curve.py Show resolved Hide resolved

sklearn/metrics/_plot/cap_curve.py Show resolved Hide resolved

lorentzenchr reviewed Apr 5, 2024

View reviewed changes

ogrisel reviewed Apr 8, 2024

View reviewed changes

sklearn/metrics/_plot/cap_curve.py Show resolved Hide resolved

lorentzenchr reviewed Apr 8, 2024

View reviewed changes

sklearn/metrics/_plot/cap_curve.py Show resolved Hide resolved

JosephBARBIERDARNAL closed this by deleting the head repository Apr 27, 2024

JosephBARBIERDARNAL mentioned this pull request May 7, 2024

ENH add CAP curve #28972

Open

17 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

FEA add CumulativeAccuracyDisplay #28752

FEA add CumulativeAccuracyDisplay #28752

JosephBARBIERDARNAL commented Apr 2, 2024

github-actions bot commented Apr 2, 2024 •

edited

Loading

glemaitre left a comment

ogrisel left a comment

ogrisel Apr 2, 2024 •

edited

Loading

glemaitre Apr 2, 2024

JosephBARBIERDARNAL Apr 3, 2024

ogrisel Apr 4, 2024

JosephBARBIERDARNAL commented Apr 3, 2024

ogrisel commented Apr 4, 2024 •

edited

Loading

ogrisel commented Apr 4, 2024 •

edited

Loading

ogrisel commented Apr 4, 2024

ogrisel left a comment

ogrisel Apr 4, 2024

ogrisel Apr 5, 2024

JosephBARBIERDARNAL commented Apr 4, 2024

ogrisel commented Apr 5, 2024

lorentzenchr commented Apr 5, 2024

ogrisel commented Apr 5, 2024

ogrisel left a comment

ogrisel Apr 5, 2024

JosephBARBIERDARNAL Jan 11, 2025

JosephBARBIERDARNAL Jan 11, 2025

lorentzenchr left a comment

lorentzenchr Apr 5, 2024

ogrisel Apr 8, 2024

lorentzenchr commented Apr 5, 2024

ogrisel commented Apr 8, 2024

ogrisel left a comment

lorentzenchr commented Apr 8, 2024

lorentzenchr commented Apr 8, 2024

lorentzenchr commented Apr 28, 2024

JosephBARBIERDARNAL commented Apr 28, 2024

JosephBARBIERDARNAL commented May 7, 2024

	def lorenz_curve(y_true, y_pred, exposure):
	y_true, y_pred = np.asarray(y_true), np.asarray(y_pred)
	exposure = np.asarray(exposure)

	# order samples by increasing predicted risk:
	ranking = np.argsort(y_pred)
	ranked_frequencies = y_true[ranking]
	ranked_exposure = exposure[ranking]
	cumulated_claims = np.cumsum(ranked_frequencies * ranked_exposure)
	cumulated_claims /= cumulated_claims[-1]
	cumulated_exposure = np.cumsum(ranked_exposure)
	cumulated_exposure /= cumulated_exposure[-1]
	return cumulated_exposure, cumulated_claims

		This is also known as a Gain or Lift Curve for classification, and a Lorenz
		curve for regression with a positively valued target.

FEA add CumulativeAccuracyDisplay #28752

FEA add CumulativeAccuracyDisplay #28752

Conversation

JosephBARBIERDARNAL commented Apr 2, 2024

Reference Issue

What does this implement/fix?

other

github-actions bot commented Apr 2, 2024 • edited Loading

✔️ Linting Passed

glemaitre left a comment

Choose a reason for hiding this comment

ogrisel left a comment

Choose a reason for hiding this comment

ogrisel Apr 2, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

JosephBARBIERDARNAL commented Apr 3, 2024

ogrisel commented Apr 4, 2024 • edited Loading

ogrisel commented Apr 4, 2024 • edited Loading

ogrisel commented Apr 4, 2024

ogrisel left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

JosephBARBIERDARNAL commented Apr 4, 2024

ogrisel commented Apr 5, 2024

lorentzenchr commented Apr 5, 2024

ogrisel commented Apr 5, 2024

ogrisel left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

lorentzenchr left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

lorentzenchr commented Apr 5, 2024

ogrisel commented Apr 8, 2024

ogrisel left a comment

Choose a reason for hiding this comment

lorentzenchr commented Apr 8, 2024

lorentzenchr commented Apr 8, 2024

lorentzenchr commented Apr 28, 2024

JosephBARBIERDARNAL commented Apr 28, 2024

JosephBARBIERDARNAL commented May 7, 2024

github-actions bot commented Apr 2, 2024 •

edited

Loading

ogrisel Apr 2, 2024 •

edited

Loading

ogrisel commented Apr 4, 2024 •

edited

Loading

ogrisel commented Apr 4, 2024 •

edited

Loading