oxits

A high-performance time series classification and transformation library for Rust, validated against pyts.

Features

Preprocessing

• StandardScaler — zero-mean, unit-variance normalization
• MinMaxScaler — scale to arbitrary range
• MaxAbsScaler — scale by maximum absolute value
• RobustScaler — median/IQR-based scaling
• KBinsDiscretizer — binning with normal, uniform, and quantile strategies
• PowerTransform — Box-Cox and Yeo-Johnson transforms
• QuantileTransform — uniform or normal output distribution
• Imputer — fill NaN values (nearest, previous, next, linear)

Approximation

• PAA — Piecewise Aggregate Approximation
• SAX — Symbolic Aggregate Approximation
• DFT — Discrete Fourier Transform coefficients
• SFA — Symbolic Fourier Approximation (DFT → MCB discretization, ANOVA feature selection)

Metrics

• DTW — Dynamic Time Warping (classic, Sakoe-Chiba band, Itakura parallelogram, multiscale, fast)
• Lower bounds — LB_Kim, LB_Keogh, LB_Improved, LB_Yi
• BOSS metric — histogram intersection distance

Image Transforms

• GASF — Gramian Angular Summation Field
• GADF — Gramian Angular Difference Field
• MTF — Markov Transition Field
• RecurrencePlot — recurrence plot with time-delay embedding

Decomposition

• SSA — Singular Spectrum Analysis with automatic trend/seasonal/residual grouping

Transformation

• BOSS — Bag of SFA Symbols
• ROCKET — Random Convolutional Kernel Transform
• BagOfPatterns — sliding-window SAX bag of words with TF-IDF
• ShapeletTransform — shapelet-based feature extraction
• WEASEL — Word ExtrAction for time SEries cLassification

Classification

• KNN — k-nearest neighbors with pluggable distance metrics
• BOSSVS — BOSS in Vector Space (TF-IDF cosine similarity)
• SAXVSM — SAX-VSM classifier
• TimeSeriesForest — interval-based random forest
• TSBF — Time Series Bag of Features
• LearningShapelets — gradient-descent shapelet learning

Multivariate

• JointRecurrencePlot — joint recurrence plots for multivariate time series
• Multivariate wrapper — apply univariate transforms/classifiers per channel

Datasets

• UCR Archive — fetch datasets from the UCR Time Series Archive
• Synthetic generators — Cylinder-Bell-Funnel (CBF) dataset
• Built-in — GunPoint, Coffee synthetic datasets

Infrastructure

• Parallel computation — feature-gated Rayon parallelism across all modules
• Core traits — Transformer, FittableTransformer, Classifier, DistanceMetric
• SIMD autovectorization — AVX2 on x86-64 via target-cpu=native

Performance

See PERFORMANCE.md for detailed benchmark tables against pyts across all algorithms.

Benchmarked on Intel i9-13900H (P25 of 51 runs, 5 warmup):

Installation

Add to your Cargo.toml:

[dependencies]
oxits = "0.1"

Quick Start

Stateless Transform (StandardScaler)

use oxits::preprocessing::scaler::{StandardScaler, StandardScalerConfig};
use oxits::Transformer;

let config = StandardScalerConfig::new();
let x = vec![vec![1.0, 2.0, 3.0, 4.0, 5.0]];
let scaled = StandardScaler::transform(&config, &x);

Stateful Transform (SFA)

use oxits::approximation::sfa::{Sfa, SfaConfig};
use oxits::FittableTransformer;

let config = SfaConfig { n_coefs: Some(4), n_bins: 4, ..SfaConfig::new() };
let x = vec![
    vec![0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0],
    vec![7.0, 6.0, 5.0, 4.0, 3.0, 2.0, 1.0, 0.0],
];
let fitted = Sfa::fit(&config, &x, None);
let result = Sfa::transform(&fitted, &x);

Classification (BOSSVS)

use oxits::classification::bossvs::{Bossvs, BossvsConfig};

let config = BossvsConfig::new(4);
let x_train = vec![
    vec![0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0],
    vec![7.0, 6.0, 5.0, 4.0, 3.0, 2.0, 1.0, 0.0],
];
let y_train = vec!["A".to_string(), "B".to_string()];

let fitted = Bossvs::fit(&config, &x_train, &y_train);
let predictions = Bossvs::predict(&fitted, &x_train);

Distance Metrics (DTW)

use oxits::metrics::dtw::dtw_classic;

let a = vec![1.0, 2.0, 3.0, 4.0, 5.0];
let b = vec![1.0, 2.5, 3.5, 4.0, 5.0];
let distance = dtw_classic(&a, &b);

Image Transform (GASF)

use oxits::image::gaf::{Gaf, GafConfig};
use oxits::{GafMethod, Transformer};

let config = GafConfig { method: GafMethod::Summation, image_size: None };
let x = vec![vec![0.0, 1.0, 2.0, 3.0, 4.0]];
let images = Gaf::transform(&config, &x);
// images[0] is a 5x5 Gramian Angular Summation Field

Cargo Features

# Default (parallel)
cargo build --release

# All features
cargo build --release --all-features

# No parallelism
cargo build --release --no-default-features

Building

cargo build --release
cargo test --all-features

For best performance, ensure .cargo/config.toml targets your CPU:

[target.x86_64-unknown-linux-gnu]
rustflags = ["-C", "target-cpu=native"]

Validation

All modules are validated against pyts via golden integration tests. Each test loads reference data generated by pyts and compares output at epsilon < 1e-6:

cargo test --all-features         # run all tests (229 total)
cargo test --test golden_image    # image module golden tests
cargo test --test golden_metrics  # DTW golden tests

To regenerate golden data:

cd test_harness
pip install pyts numpy scikit-learn
python generate_golden_data.py

Dependencies

• realfft — FFT for DFT, SFA, and SSA periodograms
• rayon — parallel computation (optional)
• rand / rand_distr — random kernels for ROCKET
• nalgebra — SVD for SSA decomposition (optional)
• ureq — HTTP client for UCR Archive (optional)

License

MIT License — see LICENSE.