Rgram
Skip to content
RgramResult schemas

Fitted attributes and result schemas

Unless noted, tables are Polars DataFrames and queries retain their input order and duplicates. Null metadata means a bin was absent; NaN predictions/bounds mean no finite estimate was available under the selected policy. Strict policies can raise before a table is returned. Never treat numerical support as accuracy.

Shared fitted state

Regressogram and KernelSmoother expose X_, y_, and sample_weight_ as independent one-dimensional training snapshots; omitted weights are ones. n_samples_in_ counts all rows, n_features_in_ is one, and sample_weight_provided_ records whether weights were supplied. feature_names_in_ exists only for named input and is cleared by an unnamed refit. The estimator class reference lists all model-specific fitted attributes.

data_summary_ has these keys:

Key Meaning
n_rows_received, n_rows_retained Training row counts; equal after successful validation.
rows_sorted, rows_dropped, values_imputed False, 0, 0 respectively.
x_dtype, y_dtype Original snapshot dtypes as strings.
computation_dtype "float64".
sample_weight_provided Whether observation weights were supplied.
weights_modified_in_snapshot False; computational normalization does not alter stored weights.
Regressogram: binning, n_bins_requested, n_bins_fitted, cv_used, extrapolation Requested bin strategy/option, fitted count, internal CV flag, and range policy.
Smoother: algorithm, weight_scale, bandwidth, extrapolation Fit-time computation path, numerical weight scale, effective fitted bandwidth, and range policy.

requested_n_bins_ is available for width/dist binning only and is the nominal count after rule/cap/sample/constant-feature adjustment, before boundary merging. selected_n_bins_ stores the internal CV winner or the original n_bins setting. For integer/exact grouping, n_bins_ counts unique groups; bin_edges_ is empty and bin_width_ is None. bins_ contains occupied groups, so its height need not equal the nominal cell count.

Regressogram bins

Column in bins_ Meaning
rgram_bin Integer cell index for dist/width, truncated integer for int, feature value for none.
y_pred_rgram Configured response aggregation.
y_pred_rgram_lci, y_pred_rgram_uci Present only when descriptive ci reductions are configured.
n_samples All observations in the cell, including zero-weight rows.
weight_sum Sum of supplied weights, or row count with omitted weights.
n_positive_weight Number of rows with positive weight.
x_min_observed, x_max_observed Actual observed feature range in the cell.
bin_left, bin_right Nominal cell limits for dist/width only; equality conventions are in the parameter reference.

Rows follow first-observed cell order, not sorted boundary order.

Prediction diagnostics

Regressogram predict_diagnostics returns:

Column Meaning
x Original query coordinate.
prediction Bin aggregation at the query.
summary_lower, summary_upper Present only with configured ci: descriptive bin endpoints, distinct from bootstrap confidence bounds.
rgram_bin Assigned group/cell, including edge assignment where applicable.
n_samples, weight_sum, n_positive_weight Information from the assigned occupied bin, or null if absent.
supported Whether the prediction is finite.
in_training_range Whether the original query lies between observed feature extrema, inclusively.

Kernel predict_diagnostics and predict_grid return:

Column Meaning
x, prediction, supported, in_training_range Same meanings as above.
n_neighbors Count of positive numerical kernel/observation weights; zero for unsupported queries.
effective_n Inverse sum of squared normalized nonnegative weights; zero for unsupported queries. A concentration measure, not degrees of freedom.
local_linear_fallback Whether this query used a local constant because its slope was unresolved.

Residual diagnostics

regression_diagnostics(X, y) returns x, observed, prediction, residual, and supported. Residual is observed minus predicted, with NaN when unsupported. It accepts training or held-out aligned pairs. No row is removed because its prediction is unsupported.

Bootstrap intervals

predict_interval returns:

Column Meaning
x Original query.
prediction Original fitted curve estimate.
lower, upper Pointwise bootstrap bounds; NaN without adequate draw support.
n_valid Number of finite bootstrap predictions for this query.
bootstrap_coverage n_valid / n_resamples, not nominal confidence coverage.

Intervals require at least two valid draws and the configured valid fraction. These bounds are distinct from regressogram descriptive ci columns.

Search results

CoverageSearchCV.cv_results_ contains one entry per candidate:

Key Meaning
params List of candidate parameter dictionaries.
mean_test_mse Pooled squared error divided by the number of finite validation predictions, not an unweighted average of fold means. Infinity if none are finite.
coverage Pooled supported validation fraction.
min_fold_coverage Lowest supported fraction among folds.
eligible Every fold meets min_coverage and pooled MSE is finite.
selection_loss MSE for eligible candidates, infinity otherwise.
n_validation, n_supported Total validation occurrences and supported occurrences; repeated validation membership counts repeatedly.
split{k}_coverage Supported fraction in fold k.
split{k}_test_mse MSE over supported rows in fold k; infinity when none are supported.

All entries except params are NumPy arrays. cv_splits_ stores copied train/test indices, and n_splits_ stores the fold count. best_index_, best_params_, best_score_ (negative selection loss), and best_estimator_ describe the refitted winner. An all-ineligible fit keeps results for inspection but does not produce a fitted predictor. Internal Regressogram(n_bins="cv") exposes the same results dictionary and its selected bin count, but not a complete search object.