Skip to main content
Understand why Nori made a prediction: which features drove it, how they interact, and how the model responds across a feature’s range. Because NoriRegressor is a scikit-learn estimator, it plugs directly into shapiq — a fast SHAP implementation with native Shapley-interaction support — and the rest of the sklearn interpretability ecosystem. The helpers in synthefy_nori.interpretability are thin convenience wrappers, so you can also use the underlying tools directly. Runnable example

​
Install

pip install "synthefy-nori[interpretability]"   # pulls in shapiq
Nori is regression-only, so every method below explains the regression predictive mean.

​
SHAP / Shapley values & interactions

get_nori_imputation_explainer builds a shapiq.TabularExplainer that removes features by imputation against a background set. The training context stays fixed across coalitions, so each coalition is a single predict call — the cost is set by budget. 
from synthefy_nori import NoriRegressor
from synthefy_nori.interpretability.shapiq import get_nori_imputation_explainer

model = NoriRegressor().fit(X_train, y_train)

# index="k-SII", max_order=2 captures pairwise interactions;
# use index="SV", max_order=1 for plain Shapley values.
explainer = get_nori_imputation_explainer(model, X_train, index="k-SII", max_order=2)
sv = explainer.explain(X_test[:1], budget=128)
print(sv)
sv.plot_waterfall()   # additive contribution waterfall
SHAP waterfall plot for a single Nori prediction
Key parameters of get_nori_imputation_explainer:
  • index — "SV" for plain Shapley values, or "k-SII" for k-Shapley interactions (max_order=1 reduces it to standard Shapley values).
  • max_order — maximum interaction order (1 = single-feature attributions, 2 = pairwise interactions).
  • imputer — "baseline" (one forward per coalition; recommended) or "marginal" / "conditional" (multi-sample, much slower).
Budget trades accuracy for cost. Start at 128 and raise only if the explanation looks noisy: <10 features → 64–128; 10–20 → 128–512; 20+ → 512–2048.
With index="k-SII", max_order=2 the explainer also surfaces pairwise interactions, which you can view as a network: 
kv = explainer.explain(X_test[:1], budget=256)
kv.plot_network()   # interaction graph
SHAP k-SII interaction network

​
Partial dependence / ICE

A global view of how the prediction shifts across a feature’s range: 
from synthefy_nori.interpretability.pdp import partial_dependence_plots

# 1-D effects for features 0 and 2; use kind="individual" or "both" for ICE.
partial_dependence_plots(model, X_test, features=[0, 2], kind="average")
Pass (i, j) tuples in features for 2-D interaction surfaces. Returns an sklearn PartialDependenceDisplay.
Partial dependence plots for bmi and bp

​
Feature selection

Find a minimal feature subset that preserves cross-validated performance (sequential selection): 
from synthefy_nori.interpretability.feature_selection import feature_selection

res = feature_selection(model, X_train, y_train, n_features_to_select=5, cv=3)
print(res.selected_indices, res.selected_score_mean)
n_features_to_select accepts an int, a fraction, or "auto" (with tol). The result also reports baseline-vs-selected CV scores.
Sequential selection re-fits Nori in-context on every CV split, so keep it to a few thousand samples and a modest feature count.

​
Notes

  • These run on the local Python package (synthefy-nori), not the hosted API.
  • TabPFN’s fast Shapley path reuses a KV-cache across coalitions; Nori’s public package runs one forward per coalition — correct and budget-controlled, just not cache-accelerated.