Synthetic data: it comes from a known formula, so you can check whether the explanation gets it right (see the help).
600 rows · classification: the probability of «1»
It gets 84% right on rows it has not seen (92.4% on the training rows).
The model says:61.3%(actual: 1)
SHAP (exact, 1280 predictions)
From the average prediction to this row's, variable by variable:
LIME
A line that imitates the model near this row. Weight of each variable (green pushes up):
Fidelity (R² of the line near the row): 0.45, with 600 made-up rows.
SHAP and LIME rank the variables almost the same (ρ = 1).
What matters most overall
The average |SHAP| of each variable over 30 test rows.
In this data, «dia_nacimiento» has no influence at all in the formula that generated it: if its bar comes out almost empty, the explanation got it right.