LOWESS Batch Smoothing — Lowess • rfastlowess

Create a stateful LOWESS model for batch smoothing. This is the default mode: it processes the entire dataset at once and supports every feature (confidence/prediction intervals, cross-validation, GPU backend).

Usage

Lowess(
    fraction = 0.67,
    ...,
    iterations = 3L,
    weight_function = "tricube",
    robustness_method = "bisquare",
    delta = NULL,
    zero_weight_fallback = "use_local_mean",
    boundary_policy = "extend",
    scaling_method = "mad",
    auto_converge = NULL,
    missing = "error",
    parallel = TRUE,
    backend = "cpu",
    outputs = NULL,
    intervals = NULL,
    cv = NULL,
    seed = NULL,
    retain_model = FALSE
)

Arguments

fraction: Smoothing fraction, greater than 0 and up to 1. Default: 0.67.
...: Not used; forces all subsequent arguments to be named.
iterations: Number of robustness iterations, between 0 and 1000 (inclusive). Default: 3.
weight_function: Kernel weight function. One of "tricube" (default), "gaussian", "uniform" (alias: "boxcar"), "cosine", "epanechnikov", "biweight" (alias: "bisquare"), or "triangle" (alias: "triangular").
robustness_method: Outlier downweighting method: "bisquare" (default; alias: "biweight"), "huber", or "talwar".
delta: Interpolation distance threshold, as a non-negative fraction of the x range; points within delta of each other on x share the same local fit. NULL (default) sets it automatically to 1/100th of the x range.
zero_weight_fallback: Fallback policy when all robustness weights drop to zero: "use_local_mean" (default; aliases: "local_mean", "mean"), "return_original" (alias: "original"), or "return_none" (alias: "none").
boundary_policy: Boundary handling strategy: "extend" (default; alias: "pad"), "reflect" (alias: "mirror"), "zero", or "noboundary" (alias: "none").
scaling_method: Residual scale estimation for robustness weights: "mad" (default; alias: "median_absolute_deviation"), "mar" (alias: "median_absolute_residual"), or "mean" (alias: "mean_absolute_residual").
auto_converge: Convergence tolerance for early stopping of robustness iterations. NULL (default) disables early stopping.
missing: Policy for non-finite (NaN/Infinity) values in input data: "error" (default) raises an error, "drop" silently removes affected observations before fitting.
parallel: Logical; enable parallel processing. Default: TRUE.
backend: Execution backend: "cpu" (default) or "gpu". GPU support requires the package to be built locally with WITH_GPU=1 (see bindings/r/Makefile) and a Vulkan/Metal/DX12-capable GPU driver; not available in released CRAN/Bioconductor binaries.
outputs: Character vector selecting optional output components: "se" (standard errors), "diagnostics", "residuals", "weights" (robustness weights), "derivative", and/or "sorted". NULL (default) returns only the core result (x, y, and fit metadata).
intervals: Interval options, created with intervals_opts (or a named list with any of confidence, prediction, bootstrap): e.g. intervals = intervals_opts(confidence = 0.90, prediction = 0.99, bootstrap = 200). Confidence and prediction coverage levels are independent and may differ. NULL (default) disables intervals.
cv: Cross-validation options, created with cv_opts: e.g. cv = cv_opts(method = "kfold", k = 5, fractions = c(0.2, 0.3, 0.5)). NULL (default) disables cross-validation.
seed: Non-negative whole-number seed shared by cross-validation and bootstrap resampling for reproducible results. When cross-validation is unavailable, it controls bootstrap resampling only. NULL (default) uses a random seed.
retain_model: Logical; if TRUE, retain the fitted model's training data, enabling predict.Lowess for out-of-sample prediction. Default: FALSE.

Value

A Lowess object.

Details

Best suited when the dataset fits in memory and you need intervals, cross-validation, or diagnostics. For datasets that don't fit in memory or arrive in chunks, see StreamingLowess; for point-by-point real-time data, see OnlineLowess. When "diagnostics" is selected, residual_sd is the robust residual scale estimate (1.4826 * MAD).

fraction is the most important parameter: it controls the size of the local neighbourhood used at each point.

Range	Effect	Use case
0.1-0.3	Fine detail	Rapidly changing signals
0.3-0.5	Balanced	General purpose
0.5-0.7	Heavy smoothing	Noisy data
0.7-1.0	Very smooth	Trend extraction

Examples

x <- seq(0, 10, length.out = 100)
y <- sin(x) + rnorm(100, 0, 0.1)
model <- Lowess(fraction = 0.2)
result <- fit(model, x, y)
plot(x, y)
lines(x, result$y, col = "red")