Added ability to create a bar plot of feature importances #133
Conversation
Codecov Report
@@ Coverage Diff @@
## master #133 +/- ##
==========================================
+ Coverage 97.3% 97.32% +0.01%
==========================================
Files 95 95
Lines 2712 2729 +17
==========================================
+ Hits 2639 2656 +17
Misses 73 73
Continue to review full report at Codecov.
|
|
how does this handle features that weren't selected in a feature selection? is it still showing them in the x-axis? |
It only shows features that were selected for that particular pipeline. Features that were dropped in this pipeline are not shown. |
|
Actually, it looks like this functionality has already existed as feature_importances_plot() in evalml/models/render.py. Does this plot need any changes @kmax12 @jeremyliweishih?
|
evalml/pipelines/pipeline_base.py
Outdated
| def plot_feature_importances(self): | ||
| feat_imp = self.feature_importances | ||
| feat_imp['importance'] = abs(feat_imp['importance']) | ||
| feat_imp = feat_imp.iloc[::-1] |
There was a problem hiding this comment.
can you just add a comment about why we need to reverse the list? i would think feature importance already returned it in the right order
evalml/pipelines/pipeline_base.py
Outdated
| @@ -252,3 +253,28 @@ def feature_importances(self): | |||
| importances.sort(key=lambda x: -abs(x[1])) | |||
| df = pd.DataFrame(importances, columns=["feature", "importance"]) | |||
| return df | |||
|
|
|||
| def plot_feature_importances(self): | |||
There was a problem hiding this comment.
we should wait for #169 to get merge in and then change the api to pipeline.plot.feature_importances
There was a problem hiding this comment.
i see this might be a bit tricky since the feature importances are store on the pipeline object rather than results dictionary.
There was a problem hiding this comment.
I currently have it so that the plot class takes in a data dictionary that can be updated to store whatever (aka more than just results) for this purpose. I wonder if there is a better implementation?
There was a problem hiding this comment.
hmm. one though i have is that we could initialize the class every time .plot is accessed. When we initialize it, we just pass the pipeline object to the constructor. that would also also use to override __call__ so that pipeline.plot() rendered the pipeline graphic in #211
|
This function is called by using |
|
@christopherbunn I think the naming is good! |
|
Naming sounds good to me! (Honestly, might be worth reducing generate_roc() to just roc() but if anything we can address that in a future PR :)) |
| @@ -75,3 +75,32 @@ def __call__(self, to_file=None): | |||
| graph.render(output_path, cleanup=True) | |||
|
|
|||
| return graph | |||
|
|
|||
| def feature_importances(self): | |||
| import plotly.graph_objects as go | |||
There was a problem hiding this comment.
why do we import this inside of the function?
christopherbunn
dismissed
jeremyliweishih’s stale review
Addressed conf.py issues
Resolves #24