Add multicollinearity data check - #1515
Conversation
Codecov Report
@@ Coverage Diff @@
## main #1515 +/- ##
=========================================
+ Coverage 100.0% 100.0% +0.1%
=========================================
Files 234 236 +2
Lines 16827 16881 +54
=========================================
+ Hits 16819 16873 +54
Misses 8 8
Continue to review full report at Codecov.
|
|
|
||
| Arguments: | ||
| X (ww.DataTable, pd.DataFrame, np.ndarray): The input features to check | ||
| threshold (float): The probability threshold to be considered an ID column. Defaults to 1.0 |
There was a problem hiding this comment.
Not related but removing extraneous argument in docstring :d
| } | ||
|
|
||
|
|
||
| def test_multicollinearity_data_check_input_formats(): |
There was a problem hiding this comment.
Would have liked adding ww test here but made more sense to reuse data from test_multicollinearity_nonnumeric_cols
…o 811_multicollinearity
bchen1116
left a comment
bchen1116
left a comment
There was a problem hiding this comment.
LGTM! Might be nice to add a np.array as a test to see whether the output columns are understandable, but not necessary.
jeremyliweishih
left a comment
jeremyliweishih
left a comment
There was a problem hiding this comment.
Current implementation looks good - just a couple questions,
-
Is this added to AutoML by default? I'm a bit worried about performance across larger datasets but maybe the WW team would know more about it.
-
Did we go with WW's implementation for convenience? I remember us looking at a VIF implementation in the past. I tried researching but couldn't find if there was a difference in using NMIF vs. just correlation as well.
I think all user facing things should be "multicollinear" but usage of correlation or correlated in the code should be fine! |
|
@jeremyliweishih Yup, currently not adding to default data checks because of what you mentioned, but open to discussing if we should or not. Would be nice to get some sort of clustering/faster algo first before doing so. And WW's normalized mutual information method handles categoricals, which is nice (and I believe something VIF doesn't) :) |
freddyaboulton
left a comment
freddyaboulton
left a comment
There was a problem hiding this comment.
@angela97lin Looks good! Maybe we should benchmark this to know where we are now in terms of performance. Obviously, that shouldn't block this merge :)
dsherry
left a comment
dsherry
left a comment
There was a problem hiding this comment.
@angela97lin LGTM! Had a couple comments. We should definitely lower the default threshold.
| if mutual_info_df.empty: | ||
| return messages | ||
| above_threshold = mutual_info_df.loc[mutual_info_df['mutual_info'] >= self.threshold] | ||
| correlated_cols = [(col_1, col_2) for col_1, col_2 in zip(above_threshold['column_1'], above_threshold['column_2'])] |
There was a problem hiding this comment.
Will this remove duplicate pairs, i.e. ('col_A', 'col_B') and ('col_B', 'col_A') ? Ideally we should find a way to do that.
There was a problem hiding this comment.
Luckily, mutual_information from Woodwork handles this; that's why we only see one pair in the expected test results 😁
6 participants
Resolves #811 by adding an initial version of a multicollinearity data check.
I switch between "correlation" and "multicollinear" because mutual information right now is technically just checking for correlation but open to thoughts about consistency 🤷