Examples¶

Explore one relationship or diagnostic at a time. Each example has its own page and a large 10 × 6.6 inch figure, saved at 180 DPI with readable labels and legends. Click a plot to inspect the original image at full resolution.

The examples use reproducible synthetic data, so some plots can show the known generating relationship. Real datasets usually do not supply that ground truth. See theory and suitability before interpreting a fitted curve.

Start with your analysis question¶

For a complete predictive workflow, start with demand from temperature: train/test splitting, pipeline preprocessing, bandwidth selection, and verified output.

• A noisy curve needs a summary: compare bin counts or bandwidths.
• The mean may hide the response distribution: compare response summaries.
• Selected groups may behave differently: fit subgroup curves and inspect their support.
• A fitted value needs explaining: review observation influence or missing support.
• You need to assess the fit: inspect residuals, boundary behavior, or curve intervals.

Use the same sequence on each page: identify the question, review the choices, read the plot, run the example, then decide what to change in your own analysis.

Browse the gallery¶

What each example examines¶

For runnable recipes using your own callbacks, see custom aggregations and smoothers: native Polars reductions, NumPy statistics, weighted reducers, descriptive endpoints, and custom kernels.

Every page describes the plotted items and their effects, explains how to read the result, identifies useful applications and limits, and suggests a concrete follow-up.

Run one plot at a time¶

python examples/plot_relationships.py --example bandwidth

Each page includes its own command and the relevant plotting function. With no --example argument, the script displays each figure separately; close the current window to continue to the next one. Install Matplotlib through python -m pip install 'rgram[plot]' first.

To reproduce all documentation figures without a desktop:

python examples/plot_relationships.py --output-dir docs/_static/gallery

Download the complete gallery script. Figures are generated from the same functions shown on the example pages.

Interactive kernel explorer¶

uv run python examples/kernel_explorer.py

The explorer provides controls for kernel, bandwidth, query location, local constant/linear fitting, binning, and aggregation. It displays local influence weights, support counts, effective sample size, and fallback flags. Bootstrap intervals are calculated only after pressing the interval button. It requires a desktop backend; GitHub Pages cannot run Python widgets. Download the explorer script and run it locally.

Exact neighbor and brute-force paths¶

uv run python examples/benchmark_neighbors.py

This script supplies already sorted features and a compact tricube kernel, then times the automatic support-window path against bounded brute-force prediction. It verifies numerical agreement. Results depend on hardware, sample size, query count, and bandwidth; they are not a universal speed claim. Download the benchmark script.

Pipelines, selection, and weighted fitting¶

The scikit-learn guide contains executable examples for fitting, cloning, preprocessing pipelines, GridSearchCV, selecting one feature from a wider table, and step-prefixed fitting weights. Custom callback examples are in the aggregation API, and coverage-aware selection is shown in the search reference.

Code structure and shared helpers¶

The gallery keeps data generation, estimation, and interpretation inside nine small example functions, one per statistical question. Each returns one figure without saving files or opening windows. main() selects the requested examples and owns display, export, and cleanup. This makes the same functions usable in headless checks and in an interactive session.

The shared helpers keep presentation consistent rather than concealing analysis: _create_axes makes one large, readable panel; _plot_observations displays original responses; _add_legend labels comparisons. The model helper below explicitly selects a Gaussian kernel and a manual bandwidth measured in feature units:

def _gaussian_smoother(bandwidth: float, **kwargs) -> KernelSmoother:
    """Keep kernel and bandwidth units explicit; kwargs select other model options."""
    return KernelSmoother(
        kernel="gaussian", bandwidth="manual", bandwidth_value=bandwidth, **kwargs
    )

Each example fixes a random seed when generating noisy data and holds irrelevant choices constant during comparisons. Sorted feature construction is explicit; Rgram itself does not sort observations. Query grids control where a fit is drawn, not how many observations it learns from. Inline comments identify intentional outliers, missing support, uncertainty assumptions, and the meaning of diagnostics.