Shap: LightGBM classifier can have negative base value?

[henrique]

Could the base values be negative on classification? Also the output on shap.force_plot doesn't match what I get from lgb.predict_proba....
I'd be happy to share the kernel or write a simpler test case if that helps.

Kaggle has some old version of shap but I get the same locally with 0.21.0

[slundberg]

It is important to keep in mind that LightGBM trees are built on the log-odds scale and then just transformed to probabilities for predict_proba. So SHAP values are also in log odds units. A negative base value means you are more likely class 0 than 1, and the sum will equal the log-odds output of the model not the transformed probability after the logistic function. Hope that is helpful!

[henrique]

Thanks heaps for the clarification Scott!
Lame me, I can easily reproduce the lgb output by using the inverse logit of the shap_values' sum (+base value)!

Amazing work btw! Thanks for sharing this lib and the research behind it.

[cbeauhilton]

@henrique Were you able to generate force plots with the probabilities instead of log odds with lgb? Would you mind sharing code if so?

[henrique]

The kernel is here https://www.kaggle.com/hmendonca/lightgbm-predictions-explained-with-shap-0-796
But it uses the logit outputs. I think you could change the code to sigmoid the values before plotting them, bringing it to a linear 0-1 scale but it doesn't really matter that much if you just want to see the contributions, does it?

[cbeauhilton]

Thanks @henrique , that was very useful!

[slundberg]

I should note that you can directly go to probability space now by using TreeExplainer's feature_dependence="independent" and model_output="probability" options. You can also transform the x-axis labels of the force plot from log odds to probabilities (while leaving the bar widths still in log odds) by using the link="logit" option of force_plot.

[SaadAhmed96]

It is important to keep in mind that LightGBM trees are built on the log-odds scale and then just transformed to probabilities for predict_proba. So SHAP values are also in log odds units. A negative base value means you are more likely class 0 than 1, and the sum will equal the log-odds output of the model not the transformed probability after the logistic function. Hope that is helpful!

Thanks for the explanation but what does a negative value tell when we have more than 2 classes.

[slundberg]

If you are going into a softmax then it is important to remember that the softmax is invariant to constant offsets added to the margin of all the classes. That means 0 does not need to be special, it is just the relation to other values that matters. However, if you were to subtract the mean margin from all classes then a negative value would just mean that class has a below average likelihood.

[NancyLele]

@henrique
I met the same problem. And tried all methods above. but the result is not probability

I think whether the method lightgbm.train do not have the function predict_proba

[slundberg]

@NancyLele that looks like it will produce probabilities, so perhaps share the plot you get?

[NancyLele]

@slundberg

the above picture is result. The result is 0 or 1,not the probability

[SaadAhmed96]

I think your results are in probability space. Try printing the result of predict_proba of your model for this same example. I believe they will be the same. Also, plot the force_plot for a few more examples in your dataset to make sure you understand the results.

[NancyLele]

@SaadAhmed96
The result of model.predict () is probability. The lightgbm.train() has not predict_proba function ,only has predict() function. I print the result of predict of my model,the result is 0.24. And I try more examples in my dataset .The reuslt of force_plot is 0 or 1,not the probability. So I don't know have to show probability in force_plot

[SaadAhmed96]

Sorry, should have seen it earlier! I tried it on my own binary classification dataset and the results are in probability. In your code, you have added to 3 more arguments in the TreeExplainer change it to TreeExplainer(model) only and than try plotting again.

[slundberg]

Thanks @SaadAhmed96

@NancyLele the output of lightgbm_model.predict(predict_) should be the same as shap_values.sum(1) + explainer.expected_value when using model_output='probability'. If it is not then please post a self contained example so we can debug it :)