Presenting Attributions

Intuition

An explanation with 300 numbers explains nothing. Show the few features that moved the prediction most, with their direction ("raised"/"lowered") and how much of the total effect they account for.

Precise explanation

Given attributions $\phi_j$ (e.g. from XA-01, XA-11):

• rank by $|\phi_j|$ descending (ties → feature name ascending, for determinism);
• direction: "+" if $\phi_j > 0$, "−" if $\phi_j < 0$, "0" otherwise;
• share $= |\phi_j| / \sum_i |\phi_i|$ (0 if all attributions are 0). Sorting by absolute value matters: a large negative effect is as important as a large positive one.

Worked example

names [age, income, debt], φ = [0.5, −2.0, 1.5] → ranking income (−, 0.5), debt (+, 0.375), age (+, 0.125).

Common mistakes

• Sorting by signed value (hides strong negative effects).
• Shares computed from signed sums (can divide by ~0).
• Non-deterministic order for ties.

Why this matters

Explanations are for people; clear, stable presentation is part of building trustworthy systems (Molnar's book discusses human-friendly explanations).

References

• Interpretable Machine Learning (3rd edition) — Christoph Molnar (S024)