I had a question regarding the shap interaction matrix. Suppose I have 500 samples with 2 features. Then my interaction matrix becomes (500,2,2).
I want to calculate the SHAP values of each feature and its interaction, thus x1,x2,x1-x2. Inormalize these values so that I can explain the contribution (in %) of x1,x2 and x1-x2 to the total shap values. Is this possible?
That means I decompose the 2by2 matrix for all 500 values, an average their absolute values to explain the contribution of x1,x2 and x1-x2. Is this possible?
I have a linear regression model in the form y ~ x1+x2+x1*x2. So I want to compare the coefficients with SHAP values. My SHAP values are from a gradient boosting model where I do not explicitly suggest the interactions, just the individual features as my assumption is that the ML model can capture interactions if there is enough depth in the trees.
