Merge Strategies

How overlapping chunk boundaries are reconciled in Streaming mode.

Overview

Streaming LOWESS processes data in fixed-size chunks with a configurable overlap. Points inside the overlap zone are fitted twice — once by the left chunk and once by the right chunk. The merge_strategy decides how those two estimates are combined into a single output value.

Chunk A:   [=========|=====]
Chunk B:            [=====|=========]
Overlap:            [=====]
                      ↑
                 merge_strategy
                 applied here
StrategyMethodRobustnessSpeed
"average"Simple mean of both estimatesLowFastest
"take_first"Left-chunk estimate onlyLowFastest
"take_last"Right-chunk estimate onlyLowFastest
"weighted_average"Distance-weighted meanHighModerate

Merge Strategies


Average

Takes the arithmetic mean of the left-chunk and right-chunk estimates in the overlap region. Fast and sufficient when both chunks have similar smoothing quality.

Use when: Chunks are large and the overlap region has uniform data density.

using FastLOWESS
using Random, Statistics

rng = MersenneTwister(42)
x = collect(range(0, 2π, length=100))
y = sin.(x) .+ randn(rng, 100) .* 0.3
x_chunk = x[1:50]; y_chunk = y[1:50]

model = StreamingLowess(; merge_strategy="average")
process_chunk(model, x_chunk, y_chunk)
result = finalize(model)
println("First smoothed value (average strategy): ", result.y[1])
First smoothed value (average strategy): 0.3588573878012469

Take First

Keeps only the left-chunk estimate in the overlap zone and discards the right-chunk estimate. Produces a definitive, non-revised output as soon as the right boundary of each chunk is reached.

Use when: You need final output values immediately after each chunk (no look-ahead revision); left-chunk data quality is higher.

using FastLOWESS
using Random, Statistics

rng = MersenneTwister(42)
x = collect(range(0, 2π, length=100))
y = sin.(x) .+ randn(rng, 100) .* 0.3

model = StreamingLowess(; merge_strategy="take_first")
process_chunk(model, x, y)
result = finalize(model)
println("First smoothed value (take_first strategy): ", round(result.y[1]; digits=4))
First smoothed value (take_first strategy): 0.4426

Take Last

Keeps only the right-chunk estimate in the overlap zone. The right chunk sees more of the surrounding data, so its fit can be more accurate near the left boundary of the new chunk.

Use when: Right-chunk context improves overlap quality; you are post-processing complete data rather than streaming live.

using FastLOWESS
using Random, Statistics

rng = MersenneTwister(42)
x = collect(range(0, 2π, length=100))
y = sin.(x) .+ randn(rng, 100) .* 0.3

model = StreamingLowess(; merge_strategy="take_last")
process_chunk(model, x, y)
result = finalize(model)
println("First smoothed value (take_last strategy): ", round(result.y[1]; digits=4))
First smoothed value (take_last strategy): 0.4426

Weighted Average

Assigns each overlap point a weight proportional to its proximity to the centre of its respective chunk: points near the left-chunk centre get higher left weight; points near the right-chunk centre get higher right weight. This produces the smoothest transition across chunk boundaries.

\[\hat{y} = \frac{w_L \hat{y}_L + w_R \hat{y}_R}{w_L + w_R}\]

where $w_L$ and $w_R$ are linear distance weights from the chunk centres.

Use when: Minimising boundary artefacts is more important than speed; moderate overlap (10–20 % of chunk size).

using FastLOWESS
using Random, Statistics

rng = MersenneTwister(42)
x = collect(range(0, 2π, length=100))
y = sin.(x) .+ randn(rng, 100) .* 0.3

model = StreamingLowess(;
    chunk_size=5000,
    overlap=500,
    merge_strategy="weighted_average",
)
process_chunk(model, x, y)
result = finalize(model)
println("First smoothed value (weighted_average strategy): ", round(result.y[1]; digits=4))
First smoothed value (weighted_average strategy): 0.4426

Choosing a Strategy

SituationRecommended Strategy
General purpose"weighted_average"
Maximum throughput"average"
Immediate finalised output"take_first"
Post-processing, right context better"take_last"
Minimising boundary artefacts"weighted_average"
Overlap size matters

A larger overlap gives the merge strategy more room to blend, reducing boundary artefacts regardless of the strategy chosen. A good starting point is 10 % of chunk_size.