DRAFT: dalex v1.0

[hbaniecki]

Draft: what if? Potentially v1.0 should be the next version.

Breaking changes:

• refactor modules: instance_level to predict_explanations and dataset_level to model_explanations
• move arena to 'true' lowercased folder
• add plot method and result attribute to LimeExplanation / consider adding LimeWrapper
• cord-flip cp barplots (will be consistent with ema)
• change plot titles (consistent with ema)

To resolve:

• Add or abandon pycaret Add native support for more frameworks #332 (comment: these are sklearn compatible)
• Add or abandon pyspark mllib Add native support for more frameworks #332 (comment: abandon)
• check if examples in EMA are working with dalex 1.0
• add fairness to Arena

Later:

• Finish by uploading to conda conda Python package #360

Won't be included in this release:

• support for multiclass
• predict_diagnostics

[pbiecek]

• move arena to 'true' lowercased folder in https://github.com/ModelOriented/DALEX/tree/master/python/dalex/dalex

[pbiecek]

if possible, a great feature would be the option to write a pdf version of a graph.
now it is pretty hard to polish plots in postprocessing

[hbaniecki]

For me this works:

# need pip install -U kaleido
# see https://plotly.com/python/static-image-export/

import dalex as dx
import pandas as pd
from sklearn.ensemble import AdaBoostClassifier
data = dx.datasets.load_titanic()
X = data.drop(['survived', 'embarked', 'class', 'gender'], axis=1)
y = data.survived
model = AdaBoostClassifier(random_state=0).fit(X, y)
exp = dx.Explainer(model, X, y)
pdp = exp.model_profile().plot(show=False)
pdp.write_image("temp.pdf", width=800, height=600)

[pbiecek]

and maybe this one:

• check if examples in EMA are working with dalex 1.0

[hbaniecki]

https://pypi.org/project/dalex/1.0.0/