Great question—this is a super common gotcha with SHAP beeswarm plots, and I’ve scratched my head over it too when I first started working with SHAP! The key here is understanding how the plot visualizes discrete vs. continuous features, not that some samples are missing SHAP values.

Let’s break down the main reasons:

• Visual overlap from discrete feature values
  Your male feature is a binary (discrete) feature, meaning it only has 2 possible values (0 and 1, presumably). In the beeswarm plot, all samples with the same feature value get aligned horizontally to the same x-position. When hundreds or thousands of points stack on top of each other, they look like a single dense cluster instead of individual points. On the other hand, daily_time_spent_onsite is a continuous feature—each sample has a unique (or nearly unique) value, so points spread out across the x-axis, making every individual dot visible and giving the illusion of more points.

• Default plot optimization for readability
  SHAP’s beeswarm plot automatically adjusts how it displays points to avoid clutter. For discrete features, it prioritizes showing the distribution of SHAP values at each feature level rather than every single overlapping dot. You can test this by tweaking plot parameters: try setting dot_size=1 or alpha=0.2 when calling shap.plots.beeswarm()—you’ll start seeing the individual points in the male clusters, and they’ll match the total number of samples.

To confirm all samples have SHAP values for every feature, you can check the underlying SHAP array directly. For example, if your SHAP values are stored in a variable shap_vals, run len(shap_vals[:, your_male_feature_index])—it should equal your total number of samples.

内容的提问来源于stack exchange，提问作者Parzival