A Julia package for discretizing numeric signals into bins (a.k.a. quantization).

SignalEncodings.jl provides three algorithms — uniform, quantile, and Jenks-style iterative — and supports a wide range of input layouts:

• scalar vectors (1-D signals)
• tabular data (n_samples × n_features matrices)
• time series (each cell is a 1-D array)
• images (each cell is a 2-D array)
• arbitrary tensors (each cell is an N-D array)

---

Installation

using Pkg
Pkg.add("SignalEncodings")

---

Algorithms

Config	Strategy	Key parameters
Uniform	Linearly spaced edges between min and max	nbins
Quantile	Edges at empirical quantiles	nbins, type
Jenks	Iterative break optimization minimizing within-bin deviation	nbins, maxiter, flux, errornorm

All configs share nbins, max_nobs, and rng for a uniform interface.

---

Quick start

Tabular data

using SignalEncodings

X = rand(Float32, 100, 4)   # 100 samples, 4 features

config = Uniform(; nbins=16)
X_bin, edges = bin(config, X)

config = Quantile(; nbins=16, type=:linear)
X_bin, edges = bin(config, X)

config = Jenks(; nbins=16, errornorm=:l1)
X_bin, edges = bin(config, X)

Time series

# Each cell is a 1-D time series vector
X = [rand(Float32, 50) for i in 1:30, j in 1:4]   # 30 samples, 4 channels

config = Uniform(; nbins=32)
X_bin, edges = bin(config, X)

Images

# Each cell is a 2-D image
X = [rand(Float32, 8, 8) for i in 1:20, j in 1:3]   # 20 samples, 3 channels

config = Quantile(; nbins=8)
X_bin, edges = bin(config, X)

Arbitrary tensors

# Each cell is a 3-D tensor
X = [rand(Float32, 3, 4, 4) for i in 1:10, j in 1:2]

config = Jenks(; nbins=4, errornorm=:l2)
X_bin, edges = bin(config, X)

---

Configuration reference

Uniform

Uniform(; nbins=64, max_nobs=1000, rng=Xoshiro(42))

• nbins: number of bins (2 ≤ nbins ≤ 255).
• max_nobs: per-bin sampling cap for edge estimation.
• rng: RNG for reproducible subsampling.

Quantile

Quantile(; type=:linear, nbins=64, max_nobs=1000, rng=Xoshiro(42))

• type: quantile interpolation method — one of :linear, :inverted, :average, :median, :normal, :matlab.
• nbins, max_nobs, rng: as above.

Jenks

Jenks(; nbins=64, maxiter=200, flux=0.1, fluxadjust=1.03,
        fluxadjust_bothways=true, errornorm=:l1,
        max_nobs=1000, rng=Xoshiro(42))

• maxiter: maximum optimization iterations.
• flux: initial boundary-shift ratio per iteration.
• fluxadjust: multiplicative flux adaptation factor.
• fluxadjust_bothways: allow flux to increase and decrease adaptively.
• errornorm: :l1 (sum of absolute deviations) or :l2 (sum of squared deviations).
• nbins, max_nobs, rng: as above.

---

Output

bin always returns a tuple (X_bin, edges):

• X_bin: binned indices (UInt8, 1-based). For multidimensional inputs the original item shape is preserved.
• edges: one edge vector per feature/column.

---

License

MIT License

---

About

Developed by the ACLAI Lab @ University of Ferrara.