1434 prediction explanations dataframe output #1781

freddyaboulton · 2021-02-02T20:15:32Z

Pull Request Description

After creating the pull request: in order to pass the release_notes_updated check you will need to update the "Future Release" section of docs/source/release_notes.rst to include this pull request by adding :pr:123.

codecov · 2021-02-02T20:35:13Z

Codecov Report

Merging #1781 (32cb9c8) into main (30fac41) will increase coverage by 0.1%.
The diff coverage is 100.0%.

@@            Coverage Diff            @@
##             main    #1781     +/-   ##
=========================================
+ Coverage   100.0%   100.0%   +0.1%     
=========================================
  Files         247      247             
  Lines       19573    19618     +45     
=========================================
+ Hits        19564    19609     +45     
  Misses          9        9

Impacted Files	Coverage Δ
...prediction_explanations/_report_creator_factory.py	`100.0% <100.0%> (ø)`
...tanding/prediction_explanations/_user_interface.py	`100.0% <100.0%> (ø)`
...nderstanding/prediction_explanations/explainers.py	`100.0% <100.0%> (ø)`
...s/prediction_explanations_tests/test_explainers.py	`100.0% <100.0%> (ø)`

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 30fac41...32cb9c8. Read the comment docs.

chukarsten

This looks good to me!

chukarsten · 2021-02-03T03:44:09Z

evalml/tests/model_understanding_tests/prediction_explanations_tests/test_explainers.py

@@ -399,30 +491,51 @@ def test_output_format_checked():
 @pytest.mark.parametrize("problem_type,output_format,answer,explain_predictions_answer,custom_index",


This is an epic parameterization.

chukarsten · 2021-02-03T03:44:16Z

evalml/tests/model_understanding_tests/prediction_explanations_tests/test_explainers.py

    input_features, y_true = pd.DataFrame(data=range(15)), pd.Series(range(15))
-    with pytest.raises(ValueError, match="Parameter output_format must be either text or dict. Received foo"):
+    with pytest.raises(ValueError, match="Parameter output_format must be either text, dict, or dataframe. Received foo"):
        explain_predictions_best_worst(None, input_features, y_true=y_true, output_format="foo")


I think we can probably get rid of the foo and bar tests, right? They don't add anything the "xml" one does.

chukarsten · 2021-02-03T03:45:21Z

evalml/tests/model_understanding_tests/prediction_explanations_tests/test_explainers.py

+            "quantitative_explanation": [0],
+        })
+        # Use side effect so that we always get a new copy of the dataframe
+        mock_make_table.side_effect = lambda *args, **kwargs: shap_table.copy()


This is fancy and cool. I'm going to need to go figure out what this does.

chukarsten · 2021-02-03T03:47:17Z

evalml/tests/model_understanding_tests/prediction_explanations_tests/test_explainers.py

+    elif output_format == "dataframe":
+        # Check dataframes equal without caring about column order
+        assert sorted(best_worst_report.columns.tolist()) == sorted(answer.columns.tolist())
+        pd.testing.assert_frame_equal(best_worst_report, answer[best_worst_report.columns])


Won't assert_frame_equal compare the columns as well? L597 might be redundant?

Yes but in the event best_worst_report has fewer columns than expected, the check will pass because answer[best_worst_report.columns] (needed to reorder the columns) will subset to the columns in best_worst_report.

The root cause is that I modify the index_id column at the beginning of the test to add the expected custom index and this unfortunately messes up the column order. I don't think there's a way to get assert_frame_equal to ignore the column order so this two-step check verifies the same columns are present in both and that they contain the same values. I think this is ok given how many different scenarios we're testing. The alternative would be to manually specify the answer for each test but that would be hairy.

dsherry

@freddyaboulton this is looking great!

I left some comments and questions on the implementation. Nothing blocking, but it would be great to close out those conversations soon.

dsherry · 2021-02-03T16:49:15Z

evalml/model_understanding/prediction_explanations/_user_interface.py

@@ -246,6 +253,10 @@ def make_text(self, *args, **kwargs):
    def make_dict(self, *args, **kwargs):
        """Makes the report section formatted as a dictionary."""

+    @abc.abstractmethod
+    def make_dataframe(self, *args, **kwargs):
+        """Makes the report section formatted as a dictionary."""


Do you mean " formatted as a dataframe?

dsherry · 2021-02-03T17:08:39Z

evalml/model_understanding/prediction_explanations/_user_interface.py

+                for key, value in heading.items():
+                    if key == "probabilities":
+                        for class_name, probability in value.items():
+                            shap_table[f'label_{class_name}_probability'] = probability


Can we replace this for with

labels = [f'label_{class_name}_probability' for class_name in value.keys()] shap_table[labels] = value.values()

I guess that code would rely on value.values() being ordered in the same way as value.keys(), and I don't see this set up as an OrderedDict, so maybe what you have is fine :)

Yea I think we can add another for-comprehension to make sure the order is the same but at that point the main benefit of that implementation over this one is saving an indent. So I think I'll stick to this one!

dsherry · 2021-02-03T17:09:40Z

evalml/model_understanding/prediction_explanations/_user_interface.py

+                shap_table['rank'] = heading['index']
+                shap_table['prefix'] = heading['prefix']
+            report.append(shap_table)
+        df = pd.concat(report).reset_index(drop=True)


So this code will create one df for each prediction, and then concatenate all the columns together horizontally? How do users know which columns correspond to which prediction?

To resolve: They'll use the prediction_number column we talked about above!

dsherry · 2021-02-03T17:21:23Z

evalml/model_understanding/prediction_explanations/_user_interface.py

@@ -423,3 +451,25 @@ def make_dict(self, data):
                                                                 data.input_features)["explanations"]
            report.append(section)
        return {"explanations": report}
+
+    def make_dataframe(self, data):


Is there some way we can have this implementation use _make_single_prediction_shap_table to make the tables per-prediction, and then concatenate them together? This would avoid confusion and bugs stemming from current or future differences in the implementations.

I think that's what we're doing! self.table_maker.make_dataframe(index, data.pipeline, data.input_features) calls _make_single_prediction_shap_table to make the shap table per prediction. Then we add metadata to each table (prediction probabilities, etc) and then we concat all the tables at the end.

dsherry · 2021-02-03T17:30:34Z

docs/source/user_guide/model_understanding.ipynb

-    "report = explain_predictions(pipeline=pipeline, input_features=X.iloc[[0, 4, 9]], include_shap_values=True)\n",
-    "print(report)"
+    "report = explain_predictions(pipeline=pipeline, input_features=X.iloc[[0, 4, 9]], include_shap_values=True,\n",
+    "                            output_format='dataframe')\n",


@freddyaboulton could you explain how to read the output here?

Each grouping of rows corresponds to a different prediction explanation, yes? And that's indicated by the "rank" column. What confuses me is that "rank" ranges from 0 to n-1 whereas the indices you passed to input_features are 0, 4 and 9. I'm concerned someone looking at the dataframe output will have trouble matching those row groups up with the original dataframe row that info corresponds to.

If I'm following correctly here, I suggest we either find a way to get the original dataframe index into this dataframe report, or we change our language a bit and call this "prediction number" instead of "rank". The latter option seems more appealing.

What do you think?

@dsherry You're spot on about how to read the table based on the grouping! I agree with the rename of rank to prediction_number. Adding the original dataframe index is possible but we don't do that in explain_prediction for the other output formats.

dsherry · 2021-02-03T17:33:03Z

docs/source/user_guide/model_understanding.ipynb

@@ -448,8 +448,18 @@
   "source": [
    "from evalml.model_understanding.prediction_explanations import explain_predictions\n",


@freddyaboulton I see the updates below are for the "Explaining Multiple Predictions" section. Should we also update the "Explaining Individual Predictions" section? I see that example still uses print(table)

I added a section on changing the output format for explain_prediction in the "Changing output formats" section!

freddyaboulton self-assigned this Feb 2, 2021

freddyaboulton marked this pull request as ready for review February 2, 2021 22:01

freddyaboulton requested review from dsherry, angela97lin, bchen1116, christopherbunn, jeremyliweishih, ParthivNaresh and chukarsten February 2, 2021 22:01

chukarsten approved these changes Feb 3, 2021

View reviewed changes

freddyaboulton added 7 commits February 3, 2021 11:53

Adding prototype in user guide. Will delete later.

6ba669c

Working implementation. Need tests.

dfae603

Adding unit tests.

746eacb

Adding release notes for pr 1781.

47afae8

Deleting unused method.

8743996

Passing input_data arg in tests.

b0ed5ac

Using float dtype for regression error metric for windows support.

cf35e57

freddyaboulton force-pushed the 1434-prediction-explanations-dataframe-output branch from f25c120 to cf35e57 Compare February 3, 2021 16:54

dsherry approved these changes Feb 3, 2021

View reviewed changes

dsherry reviewed Feb 3, 2021

View reviewed changes

freddyaboulton added 3 commits February 3, 2021 13:37

renaming rank to prediciton_number in explain_predictions output.

d8bf3af

Linting.

d848775

Adding tests for coverage.

32cb9c8

freddyaboulton added the enhancement An improvement to an existing feature. label Feb 4, 2021

freddyaboulton merged commit 42e01f3 into main Feb 4, 2021
2 checks passed

freddyaboulton deleted the 1434-prediction-explanations-dataframe-output branch February 4, 2021 15:12

ParthivNaresh mentioned this pull request Feb 9, 2021

Release v0.18.2 #1811

Merged

1434 prediction explanations dataframe output #1781

1434 prediction explanations dataframe output #1781

freddyaboulton commented Feb 2, 2021

codecov bot commented Feb 2, 2021 •

edited

chukarsten left a comment

chukarsten Feb 3, 2021

chukarsten Feb 3, 2021

freddyaboulton Feb 3, 2021

chukarsten Feb 3, 2021

chukarsten Feb 3, 2021

freddyaboulton Feb 3, 2021

dsherry left a comment

dsherry Feb 3, 2021

freddyaboulton Feb 3, 2021

dsherry Feb 3, 2021

freddyaboulton Feb 3, 2021

dsherry Feb 3, 2021

freddyaboulton Feb 3, 2021

dsherry Feb 3, 2021

freddyaboulton Feb 3, 2021

dsherry Feb 3, 2021

freddyaboulton Feb 3, 2021 •

edited

freddyaboulton Feb 3, 2021

dsherry Feb 3, 2021 •

edited

freddyaboulton Feb 3, 2021

		@@ -399,30 +491,51 @@ def test_output_format_checked():
		@pytest.mark.parametrize("problem_type,output_format,answer,explain_predictions_answer,custom_index",

		@@ -448,8 +448,18 @@
		"source": [
		"from evalml.model_understanding.prediction_explanations import explain_predictions\n",

1434 prediction explanations dataframe output #1781

1434 prediction explanations dataframe output #1781

Conversation

freddyaboulton commented Feb 2, 2021

Pull Request Description

codecov bot commented Feb 2, 2021 • edited

Codecov Report

chukarsten 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

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dsherry 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

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

Choose a reason for hiding this comment

Choose a reason for hiding this comment

freddyaboulton Feb 3, 2021 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dsherry Feb 3, 2021 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

codecov bot commented Feb 2, 2021 •

edited

freddyaboulton Feb 3, 2021 •

edited

dsherry Feb 3, 2021 •

edited