Add normalization option for confusion matrix #484
Conversation
|
Note: sklearn adds a normalization parameter in versions >0.22(.2?), but since our current dependencies allow for anything greater than 0.21.3 and since it was easier to just add our own normalization calculations rather than changing how |
Codecov Report
@@ Coverage Diff @@
## master #484 +/- ##
==========================================
+ Coverage 98.26% 98.26% +<.01%
==========================================
Files 104 104
Lines 3292 3352 +60
==========================================
+ Hits 3235 3294 +59
- Misses 57 58 +1
Continue to review full report at Codecov.
|
| A normalized version of the input confusion matrix. | ||
|
|
||
| """ | ||
| return conf_mat.astype('float') / conf_mat.sum(axis=1)[:, np.newaxis] |
There was a problem hiding this comment.
what's the [:, np.newaxis] for? in my testing it looks like it doesn't do anything
There was a problem hiding this comment.
My understanding is it changes broadcasting for the division. Here's an example where it makes a difference:
There was a problem hiding this comment.
hmm. i see. looks like that's need for supporting numpy arrays, but perhaps not needed for dataframes (my example).
i'm fine leaving it in to be safe, but perhaps leave a comment? This stack overflow resource was helpful for me: https://stackoverflow.com/questions/8904694/how-to-normalize-a-2-dimensional-numpy-array-in-python-less-verbose
There was a problem hiding this comment.
Oo, nice resource! Just took a closer look at the example you posted but it seems like even with dataframes, it makes a difference? The columns are flipped but the numbers outputted are also different 🤔
evalml/utils/gen_utils.py
Outdated
|
|
||
| """ | ||
| column_names = None | ||
| if isinstance(conf_mat, pd.DataFrame): |
There was a problem hiding this comment.
do you do this because of the call to np.nan_to_num(conf_mat) later? if so, I'd just end the function with
if isinstance(conf_mat, pd.DataFrame):
conf_mat = conf_mat.fillna(9)
else:
conf_mat = np.nan_to_num(conf_mat)
this would be cleaner than storing the column names
There was a problem hiding this comment.
Ah, you're right. Thanks for the suggestion, updated!
Closes #432.
Without normalization:
With normalization:
===
Since coloration is the same there, I inputted some hard-coded data to show the difference between normalization / not-normalized colors:
Without normalization:
With normalization: