Shap: some question about weird output value and base value

[Twinkle123321]

Hi,
thanks for your amazing package,I applied it to binary classification
however,I dont't know why the output value and base value are so weird
my model is LightGBM(python 3.7)
the shap_values1 is 2-d,when I input it into shap.force_plot directly,it errors
I guess it represend base value 0/1 respectively,so I chose shap_values1[1] as the first param in shap.force_plot,however,the output value was so weird,it should be range 0 from 1, could you give me some questions about it?
Thanks very much,I'm looking forward your rely!

[Twinkle123321]

[Twinkle123321]

in the link of "Census income classification with LightGBM (https://slundberg.github.io/shap/notebooks/tree_explainer/Census%20income%20classification%20with%20LightGBM.html)
I found the ouptput value in it is also negative (-8.43),how can it be explained?

[cejwlim11]

Try add option : link = 'logit'

shap.force_plot( explainer.expected_value[1], shap_values[1][0,:], link = 'logit' )

[Twinkle123321]

wow，it woks ,thank you very much!

[slundberg]

By default we explain the log-odds margin output of LightGBM models since that is a good space for additivity of effects. If you want to explain the probability output you can use model_output='probability'. Using the link option of the force plot is another option that transforms the axes values from log odds to probability for plotting.

[Twinkle123321]

Thank you for your reply，when the output value of link = 'logit' not equal to original LightGBM's probability ,I realized it may not be correct.
However,I don't find the param of model_output='probability' in shap.force_plot and shap.TreeExplainer ,could you give me some hints?
Thank you very much!

[Twinkle123321]

I find it in shap.TreeExplainer however,it show errors

assert model_output == "margin", "Only margin model_output is supported for feature_dependence="tree_path_dependent""

AssertionError: Only margin model_output is supported for feature_dependence="tree_path_dependent"

[Twinkle123321]

model_output='probability' links to feature_dependence="independent",
feature_dependence="independent" links to data is background dataset
when I set data to my train_data,it still not work,showing
AttributeError: 'TreeEnsemble' object has no attribute 'values'
the code is as follows:
explainer = shap.TreeExplainer(model_top, data = X_test_top, model_output='probability')
how can I set 'data'?

[Twinkle123321]

I solved my problem ,it seems that model should be fit again.
I still have a confusion,my dataset is unbalanced,is about class 0:class 1=1:12, how to choose background dataset?
in the notion ,says "Anywhere from 100 to 1000 random background samples are good
sizes to use.",should I create a background dataset in class 0:class 1=1:12?
Thanks again!

[slundberg]

The background dataset determines the prior expectation we have before seeing the data for the example we are explaining. So it normally makes sense to just use a random subsample of the training data. But sometimes you want to start with a different prior, like perhaps just those people who were accepted for a loan by the algorithm. That way you explain why people are different from those who got a loan, rather than why they are different from the training dataset.

[Twinkle123321]

thanks your explaination!
above mentioned
AssertionError: Only margin model_output is supported for feature_dependence="tree_path_dependent"
I find the reason is the category features leading to...