I am using the SHAP-library for computing feature Shapley values for a binary RandomForestClassifier which has naturally two outputs, 0 or 1. The forest itself consists from 100 decision tree classifiers.

I used (I assume selection happens automatically on SHAP-library based on the model) the TreeExplainer to calculate the shap values (shap_values-function) and used the layered violin plot to visualize them. I am using all default parameters in SHAP-library functions.

The layered violin plot has on the x-axis (see the link) “SHAP value (impact on model output) and on my plot, the SHAP values on the plot range from -0.4 to 0.4 with 0 on the center of the violin plot. I’ve been trying to clarify to myself what this SHAP value actually means in my case of a random forest binary classifier. Note, in my case, I am considering the SHAP value for the positive class (1).

So my question is: what does this SHAP value on the layered violin plot mean for a Scikit-learn RandomForestClassifier with binary outputs 0 or 1? Is it the log-odds? or is it the the change in the model output? But if it’s the change in model output, then what does “change in model output of -0.1 or 0.2” mean in the case where the model output is either 0 or 1?