SHAP xAI Graph scattered points

Question

I use the SHAP xAI in Python for explaining the whole test data as can be seen

.

I would like to know how are the points scattered in each line? What are the denser point areas meaning? Why are there spaces between the scattered areas in each row?

Answer 1

Y axis is feature names. Each line on scatter plot represents a feature.

X axis is SHAP values. There are equal number of points in every line: ie number of data points in your data set. Each point on line depicts SHAP value produced by this particular point. Clustering of values mean these feature [values] tend to produce similar SHAP values (due to insensitivity of output or low dispersion of feature values themselves).

The coloring of points are feature values in original units.

So putting it all together, one might state for the bottom row:

• low feature values tend to produce low SHAP(ley) values (blue datapoints have negative SHAP values)
• outcome is pretty insensitive to the feature values (because SHAP values are clustered around zero)
• sensitivity of outcome to this feature saturates/clipped at low negative values (there are no negative spikes)
• there are at least 2 datapoints with high [this particular] feature value having higher than average SHAP values.

Note, any time I say "insensitive" I mean "average marginal contribution is low over all possible coalitions".