Visualizing results and explanations

Good visualizations separate model-selection evidence from blind evaluation and distinguish global summaries from individual explanations.

Selection curves

Plot development performance against feature count for every metric used in selection:

import matplotlib.pyplot as plt

fig, axes = plt.subplots(1, 3, figsize=(12, 3.5))
for ax, metric in zip(axes, ("auc", "f1", "score")):
    plt.sca(ax)
    model.plot_performance(metric=metric)
    ax.set_title(metric.upper())
fig.tight_layout()

Illustrative backward-selection trajectory. The gold ring marks the best observed point and the dashed line marks the selected ten-feature region.

Show uncertainty or member trajectories when possible. Mark the chosen knee; do not present the best point without the path that produced it.

Boundary-counterfactual views

For classification IG that uses automatically inferred references, retrieve the already-computed boundary explanation from the attribution result:

counterfactuals = result.counterfactuals

fig, axes = plt.subplots(1, 2, figsize=(12, 5))
counterfactuals.summary_plot(ax=axes[0], max_features=12)
counterfactuals.sample_plot(
    sample_index=0,
    model_index=0,
    ax=axes[1],
    max_features=8,
)
fig.tight_layout()

Illustrative boundary-counterfactual views. The summary shows which scaled features move most on average; the sample view shows the direction and size of each selected change for one ensemble member.

The summary ranks features by mean absolute movement from observed samples to their optimized boundary points. The sample view connects observed and counterfactual values for the features that moved most. If model_index is omitted, plots average over member-specific counterfactuals; show individual members when disagreement matters. Always report feature scaling, convergence, boundary residuals, and any feasibility constraints applied after optimization.

Integrated-gradient heatmaps

A heatmap shows sample heterogeneity and co-occurring attribution patterns:

fig, ax = plt.subplots(figsize=(9, 7))
result.heatmap(
    ax=ax,
    cluster=True,
    cmap="coolwarm",
    target_strip_width=0.20,
    strip_pad=0.25,
)
ax.set_title("Integrated gradients by observation")
fig.tight_layout()

Illustrative clustered attribution profiles. Red and blue encode positive and negative integrated gradients; the side strip labels B and M classes.

A diverging color map should be centered at zero. Clustering is descriptive; cluster boundaries are not validated subtypes.

Summary plots

The summary plot combines attribution magnitude, direction, and observed feature value:

fig, ax = plt.subplots(figsize=(9, 7))
result.summary_plot(
    ax=ax,
    max_features=20,
    cmap="coolwarm",
    random_state=7,
    scatter_kwargs={"s": 28, "alpha": 0.7, "edgecolors": "none"},
)
ax.set_xlabel("Integrated-gradient attribution")
fig.tight_layout()

Illustrative summary view. Horizontal position gives attribution, color gives standardized feature value, and vertical ordering follows mean absolute attribution.

Retain sign on the horizontal axis. A bar chart of absolute means alone hides whether a feature raises or lowers outputs for different samples.

Interaction views

An interaction plot explores whether one feature’s attribution changes with a second feature:

result.interaction_plot(
    feature="Concave Points mean",
    interaction_feature="Texture worst",
    scatter_kwargs={"s": 36, "alpha": 0.75, "edgecolors": "none"},
)

This is an attribution-based interaction heuristic, not a formal statistical interaction test. Use held-out follow-up analysis for confirmatory claims.

Illustrative dependence view. Curvature or color separation can nominate an interaction for follow-up, but does not establish a statistical interaction.

Reporting checklist

Label the modeled output, class orientation, reference point, feature scale, sample cohort, and whether output came from the unified model or member set. Use the same feature names and ordering across rank tables, plots, and exported predictions.