🔅 Shapash makes Machine Learning models transparent and understandable by everyone

Description of Problem:

I want to make use of a Imputer (SimpleImputer) inside column transformer. I don't see any compatibility with Imputers.

Overview of the Solution:

Not sure how to add it yet.

Sorry but I don't see any issue with label question.

Description of Problem: Currently only Python 3.6 to 3.9 seem to be officially supported.

Overview of the Solution: Drop support for 3.6 which is EOL, add support of 3.10. Check dependencies, run tests, adapt GitHub workflow to 3.10, etc.

Code

import numpy as np
import pandas as pd

import xgboost

import shap
from shapash.explainer.smart_explainer import SmartExplainer

X,y = shap.datasets.nhanesi()
X_display,y_display = shap.datasets.nhanesi(display=True) # human readable feature values
y = np.array(y)
X = X.drop('Unnamed: 0', axis=1)

xgb_train = xgboost.DMatrix(X, label=y)

params_train = {
    "eta": 0.002,
    "max_depth": 3,
    "objective": "survival:cox",
    "subsample": 0.5,
}

model = xgboost.train(params_train, xgb_train, num_boost_round=5)

# Smart explainer creation
xpl = SmartExplainer()
xpl.compile(
    x=X,
    model=model
)

Error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-23-8abc6cc3f74d> in <module>
     28 # Smart explainer creation
     29 xpl = SmartExplainer()
---> 30 xpl.compile(
     31     x=X,
     32     model=model,

~/.local/lib/python3.8/site-packages/shapash/explainer/smart_explainer.py in compile(self, x, model, explainer, contributions, y_pred, preprocessing, postprocessing)
    192             raise ValueError("You have to specify just one of these arguments: explainer, contributions")
    193         if contributions is None:
--> 194             contributions, explainer = shap_contributions(model, self.x_init, self.check_explainer(explainer))
    195         adapt_contrib = self.adapt_contributions(contributions)
    196         self.state = self.choose_state(adapt_contrib)

~/.local/lib/python3.8/site-packages/shapash/utils/shap_backend.py in shap_contributions(model, x_df, explainer)
     55 
     56     if str(type(model)) not in list(sum((simple_tree_model,catboost_model,linear_model,svm_model),())):
---> 57         raise ValueError(
     58             """
     59             model not supported by shapash, please compute contributions

ValueError: 
            model not supported by shapash, please compute contributions
            by yourself before using shapash

Hint:

str(type(model))
"<class 'xgboost.core.Booster'>"

Python version : 3.8

Shapash version : 1.1.0 XGBoost version : 1.0.0

Operating System : Linux

Hi Team,

Is it possible to generate local explanations on the new data added to SmartPredictor object? I went through all the tutorials and I understand that when we are satisfied with the explainability results given by Shapash, we can use the SmartPredictor object for deployment. But my question is how can we get local explanation chart after deployment on the new data that is coming in? I hope my question is clear.

Thanks, Chetan

In this demo Webapp, there are some bugs when navigate : https://shapash-demo.ossbymaif.fr/ https://shapash-demo2.ossbymaif.fr/

When there are filters and you click on features in "features importance", in contribution plot, some plots do not correspond anymore to the feature and selection.

These errors are not reproducible on a local webapp, there must be a problem when deploying the exposed demos

Dear all, Many thanks for developing shapash, I have one question regarding the probability that is shown in the web app. Is it referring to the probability of the selected class or the probability of the positive class. If its the former, how someone can output the probability of the positive class.

Best regards Nikos

Hello,

I try to reproduce tutorial01-Shapash-Overview-Launch-WebApp.ipynb with my own dataset and model (RandomForestClassifier) but I have this error C extension was not built during install! when I run the .compile.

Thanks, Pauline.

Hello, I tried to build an explanation with lime backend but got the following error: My code:

from shapash.explainer.smart_explainer import SmartExplainer
import lime.lime_tabular
xpl = SmartExplainer(model=rf, backend='lime')
xpl.compile(x=X_i)

=>

File c:\Users\X\AppData\Local\Programs\Python\Python39\lib\site-packages\shapash\backend\lime_backend.py:39, in LimeBackend.run_explainer(self, x)
...
---> [39](file:///c%3A/Users/X/AppData/Local/Programs/Python/Python39/lib/site-packages/shapash/backend/lime_backend.py?line=38) explainer = lime_tabular.LimeTabularExplainer(
...

NameError: name 'lime_tabular' is not defined

I am using: Python version : 3.9.12 Lime version: 0.2.0.1

Shapash version : 2.0.1

Operating System : Windows 10

Should I import or install a particular lime version to make it work? Thanks, Thomas

Doesn't shapash support random forest for binary classification?

When I run my code, I got the below error

ValueError: 
            model not supported by shapash, please compute contributions
            by yourself before using shapash

Description of Problem:

When generating the html report, a lot of deprecation warnings are present

Overview of the Solution:

One possible solution is to add an argument to the generate_report function of the SmartExplainer class, to give the option to suppress these warnings with this snippet:

` import warnings

warnings.filterwarnings("ignore", category=DeprecationWarning) `

Python version : Python 3.8.3

Shapash version : Version: 1.4.4

Operating System : Windows 10

---------------------------------------------------------------------------
ImportError                               Traceback (most recent call last)
~\AppData\Roaming\Python\Python38\site-packages\numba\core\typeconv\typeconv.py in <module>
      3     # Numba, if it fails to import, provide some feedback
----> 4     from numba.core.typeconv import _typeconv
      5 except ImportError as e:

ImportError: DLL load failed while importing _typeconv: The specified module could not be found.

During handling of the above exception, another exception occurred:

ImportError                               Traceback (most recent call last)
<ipython-input-11-9cd225270d78> in <module>
----> 1 from shapash.explainer.smart_explainer import SmartExplainer
      2 # xpl = SmartExplainer(features_dict=house_dict)
      3 
      4 # xpl.compile(
      5 #     x=Xtest,

D:\installed_softwares\anaconda\lib\site-packages\shapash\explainer\smart_explainer.py in <module>
     14 from shapash.utils.utils import get_host_name
     15 from shapash.utils.threading import CustomThread
---> 16 from shapash.utils.shap_backend import shap_contributions, check_explainer, get_shap_interaction_values
     17 from shapash.utils.check import check_model, check_label_dict, check_ypred, check_contribution_object,\
     18     check_postprocessing, check_features_name

D:\installed_softwares\anaconda\lib\site-packages\shapash\utils\shap_backend.py in <module>
     14 import pandas as pd
     15 import numpy as np
---> 16 import shap
     17 from shapash.utils.model_synoptic import simple_tree_model, catboost_model, linear_model, svm_model
     18 

D:\installed_softwares\anaconda\lib\site-packages\shap\__init__.py in <module>
     10     warnings.warn("As of version 0.29.0 shap only supports Python 3 (not 2)!")
     11 
---> 12 from ._explanation import Explanation, Cohorts
     13 
     14 # explainers

D:\installed_softwares\anaconda\lib\site-packages\shap\_explanation.py in <module>
     10 from slicer import Slicer, Alias, Obj
     11 # from ._order import Order
---> 12 from .utils._general import OpChain
     13 
     14 # slicer confuses pylint...

D:\installed_softwares\anaconda\lib\site-packages\shap\utils\__init__.py in <module>
----> 1 from ._clustering import hclust_ordering, partition_tree, partition_tree_shuffle, delta_minimization_order, hclust
      2 from ._general import approximate_interactions, potential_interactions, sample, safe_isinstance, assert_import, record_import_error
      3 from ._general import shapley_coefficients, convert_name, format_value, ordinal_str, OpChain
      4 from ._show_progress import show_progress
      5 from ._masked_model import MaskedModel, make_masks

D:\installed_softwares\anaconda\lib\site-packages\shap\utils\_clustering.py in <module>
      2 import scipy as sp
      3 from scipy.spatial.distance import pdist
----> 4 from numba import jit
      5 import sklearn
      6 import warnings

~\AppData\Roaming\Python\Python38\site-packages\numba\__init__.py in <module>
     22 
     23 # Re-export typeof
---> 24 from numba.misc.special import (
     25     typeof, prange, pndindex, gdb, gdb_breakpoint, gdb_init,
     26     literally, literal_unroll,

~\AppData\Roaming\Python\Python38\site-packages\numba\misc\special.py in <module>
      1 import numpy as np
      2 
----> 3 from numba.core.typing.typeof import typeof
      4 from numba.core.typing.asnumbatype import as_numba_type
      5 

~\AppData\Roaming\Python\Python38\site-packages\numba\core\typing\__init__.py in <module>
----> 1 from .context import BaseContext, Context
      2 from .templates import (signature, make_concrete_template, Signature,
      3                         fold_arguments)

~\AppData\Roaming\Python\Python38\site-packages\numba\core\typing\context.py in <module>
      9 import numba
     10 from numba.core import types, errors
---> 11 from numba.core.typeconv import Conversion, rules
     12 from numba.core.typing import templates
     13 from .typeof import typeof, Purpose

~\AppData\Roaming\Python\Python38\site-packages\numba\core\typeconv\rules.py in <module>
      1 import itertools
----> 2 from .typeconv import TypeManager, TypeCastingRules
      3 from numba.core import types
      4 
      5 

~\AppData\Roaming\Python\Python38\site-packages\numba\core\typeconv\typeconv.py in <module>
     15            "possible please include the following in your error report:\n\n"
     16            "sys.executable: %s\n")
---> 17     raise ImportError(msg % (url, reportme, str(e), sys.executable))
     18 
     19 from numba.core.typeconv import castgraph, Conversion

ImportError: Numba could not be imported.
If you are seeing this message and are undertaking Numba development work, you may need to re-run:

python setup.py build_ext --inplace

(Also, please check the development set up guide https://numba.pydata.org/numba-doc/latest/developer/contributing.html.)

If you are not working on Numba development:

Please report the error message and traceback, along with a minimal reproducer
at: https://github.com/numba/numba/issues/new

If more help is needed please feel free to speak to the Numba core developers
directly at: https://gitter.im/numba/numba

Thanks in advance for your help in improving Numba!

The original error was: 'DLL load failed while importing _typeconv: The specified module could not be found.'
--------------------------------------------------------------------------------
If possible please include the following in your error report:

sys.executable: D:\installed_softwares\anaconda\python.exe

Shapash Webapp displays incorrect plot when server process > 1

In the case the webapp is exposed with processes > 1 in the uwsgi configuration, the app does not behave as expected. For instance:

• the selection of a feature triggers the display of a subset
• maximizing the feature selector plot change the selected feature
• etc.

Python version : 3.9

Shapash version : tested on 2.2, may affect other versions

Operating System : Any

How to reproduce:

app = xpl.smartapp.app
app.run_server(debug=False, host="0.0.0.0", port=8080, processes=2, threaded=False)

Description of Problem:

Shapash can use ACV as a backend to compute feature contribution. ACV has the advantage of being able to directly compute shapley values for coalition of features. In that case, their are no individual contributions for the individual features, just one for the coalition

Shapash already has a features_groups argument for the smart explainer that sum contributions of features for the group (as it should for SHAP or LIME backend). Nevertheless, the coalition property of ACV is not used in that case.

Overview of the Solution:

The feature_groups allows the user to group contributions by summing them but also to display individual contributions for the group (like in the webapp).

If we use the feature_groups argument to compute contributions for coalitions, this won't be possible for ACV. To keep the app as it is, we should add a new argument (like "acv_feature_coalition") that can be used only with acv as a backend and used in the acv_tree algorithm in acv_backend.py.

shapley_values = acvtree.shap_values(X, C=acv_feature_coalition)

Also I would suggest letting the user choose between regular shapley values computed with the acvtree.shap_values and the active shapley values computed with acvtree.shap_values_acv_adap

Description of Problem:

• PyTorch is a great multi-purpose neural network development framework.
• Deep learning algorithms are getting easier to train, understand, deploy and manage and can be used with great efficiency in enterprise contexts.
• SHAP values can be computed for neural nets and serve for model explainability purposes.

Overview of the Solution:

• Integrate the DeepLIFT algorithm (Deep SHAP), and the possibility to visualize the estimated values in the SHAPASH GUI.

Examples:

• Attribution algorithms available in the Captum Python lib
  • SHAP
  • LIME

Blockers:

• Not all explainability methods will be suitable, this is a limitation rather than a blocker

Definition of Done:

• PyTorch models trained on tabular data (regression or classification use cases) can be used with SHAPASH

First of all, thank you for the great library! Investigating shap values with it is pretty comfortable.

Description of Problem: The Shapax App does not support DataFrames with MultiIndexes. MultiIndexes are useful for querying tables coming from other structured data formats. At the xpl.run_app() step an error occurs at https://github.com/MAIF/shapash/blob/master/shapash/webapp/smart_app.py#L149.

Overview of the Solution: Use reset_index() instead of assigning the index to a column named index. If need be, store the names of the index columns for differentiating them from the remaining dataframe.

Examples:

Blockers:

Definition of Done:

I was going through the documentation and came across the below plot and its header

Here, for the local explanation id = 3, we see that SHAPASH graph plots as header as

Response: 1 - Proba = 0.2521

This is confusing. Does it mean the model predicted probability for this record was 0.7479 (obtained by subtracting 1-0.2521)?

Or is the model predicted probability of this class was 0.2521?

Which is the probability that was predicted by the model that we build?

and what does SHAPASH do here/why does it subtract from 1?