scirs2-transform

Data transformation, dimensionality reduction, and feature engineering library for machine learning in Rust, part of the SciRS2 scientific computing ecosystem.

Overview

scirs2-transform provides comprehensive data preprocessing and transformation utilities following scikit-learn's fit / transform / fit_transform API pattern. v0.3.0 significantly extends the library with UMAP, Barnes-Hut t-SNE, persistent homology / TDA, metric learning, kernel methods, optimal transport, and advanced NMF variants.

Features

Normalization and Scaling

• Min-Max scaling to [0, 1] or custom ranges
• Z-score standardization (zero mean, unit variance)
• Robust scaling (median and IQR; outlier-resistant)
• Max-absolute scaling
• L1 / L2 vector normalization
• Quantile normalization
• Reusable Normalizer with fit / transform / inverse_transform

Feature Engineering

• Polynomial features (degree 2+, with/without interaction-only mode)
• Box-Cox and Yeo-Johnson power transformations with optimal lambda estimation
• Equal-width and equal-frequency discretization (binning)
• Binarization with configurable thresholds
• Log transformations with epsilon handling
• Interaction terms, custom function transformers

Dimensionality Reduction

• PCA (Principal Component Analysis) with centering/scaling, explained variance ratio
• Truncated SVD (memory-efficient for sparse data)
• Linear Discriminant Analysis (LDA) for supervised reduction
• t-SNE with Barnes-Hut approximation (O(n log n), multicore)
• UMAP (Uniform Manifold Approximation and Projection)
• Isomap (geodesic-distance manifold learning)
• Locally Linear Embedding (LLE)
• Kernel PCA (RBF, polynomial, sigmoid kernels)
• Probabilistic PCA (PPCA)
• Factor analysis

Independent Component Analysis

• FastICA (fixed-point iteration)
• Spatial ICA
• Infomax ICA

Non-Negative Matrix Factorization (NMF) Variants

• Standard NMF (multiplicative update rules)
• Sparse NMF
• Convex NMF
• Semi-NMF
• Online NMF for streaming data

Sparse PCA and Dictionary Learning

• Sparse PCA via LASSO-based encoding
• Dictionary learning (K-SVD style)
• Sparse coding and reconstruction

Metric Learning

• Mahalanobis distance learning
• LMNN (Large Margin Nearest Neighbor)
• NCA (Neighborhood Components Analysis)
• Metric learning extensions (metric_learning_ext/)

Kernel Methods

• Kernel PCA with multiple kernels
• Deep kernel learning
• Random Fourier Features (RFF) for large-scale kernel approximation
• Orthogonal Random Features (ORF)
• Nystrom approximation

Optimal Transport

• Wasserstein distance computation
• Sinkhorn-Knopp regularized OT
• Sliced Wasserstein distance
• OT-based domain adaptation

Topological Data Analysis (TDA)

• Vietoris-Rips complex construction
• Persistent homology computation (Betti numbers, persistence diagrams)
• Persistence landscape features
• Topological feature vectorization
• Persistent diagram analysis

Archetypal Analysis

• Archetypal analysis (convex hull vertex finding)
• Simplex-based data representation

Autoencoder-Based Reduction

• Linear autoencoder as PCA surrogate
• Nonlinear autoencoder reduction

Categorical Encoding

• One-hot encoding (sparse and dense)
• Ordinal / label encoding
• Target encoding with regularization
• Binary encoding for high-cardinality features
• Unknown category handling strategies

Missing Value Imputation

• Simple imputation: mean, median, mode, constant
• KNN imputation with multiple distance metrics
• Iterative imputation (MICE algorithm)
• Missing indicator tracking

Feature Selection

• Variance threshold filter
• Recursive Feature Elimination (RFE)
• Mutual information-based selection

Pipeline API

• Sequential transformation chains
• ColumnTransformer for per-column transforms
• fit / transform / fit_transform / inverse_transform throughout

Signal Transforms (Integrated)

• Discrete Wavelet Transform (DWT): Haar, Daubechies, Symlet, Coiflet; multi-level
• 2D DWT for image decomposition (LL, LH, HL, HH subbands)
• Continuous Wavelet Transform (CWT): Morlet, Mexican Hat, Gaussian
• Wavelet Packet Transform (WPT) with best-basis selection
• Short-Time Fourier Transform (STFT) with multiple window functions
• Spectrograms (power, magnitude, dB scale)
• MFCC with mel filterbank, delta and delta-delta features
• Constant-Q Transform (CQT) for musical analysis
• Chromagram (12-bin pitch class profiles)

Multi-View Learning

• Multi-view PCA and CCA
• Consensus multi-view embedding

Online / Incremental Learning

• Incremental PCA (chunk-by-chunk update)
• Online NMF
• Online t-SNE approximations

Out-of-Core Processing

• Chunked array reader/writer for datasets larger than RAM
• Streaming normalizer with partial-fit

Structure Learning

• Covariance structure estimation
• Graphical LASSO for sparse inverse covariance

Quick Start

Add to your Cargo.toml:

[dependencies]
scirs2-transform = "0.3.0"

With SIMD and parallel features:

[dependencies]
scirs2-transform = { version = "0.3.0", features = ["parallel", "simd"] }

Normalization

use scirs2_transform::normalize::{normalize_array, NormalizationMethod, Normalizer};
use scirs2_core::ndarray::Array2;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let data = Array2::<f64>::from_shape_vec((4, 3), vec![
        1.0, 2.0, 3.0,
        4.0, 5.0, 6.0,
        7.0, 8.0, 9.0,
        10.0, 11.0, 12.0,
    ])?;

    // One-shot Z-score normalization
    let normalized = normalize_array(&data, NormalizationMethod::ZScore, 0)?;

    // Reusable normalizer (fit on train, apply to test)
    let mut scaler = Normalizer::new(NormalizationMethod::MinMax, 0);
    let train_scaled = scaler.fit_transform(&data)?;
    // let test_scaled = scaler.transform(&test_data)?;

    println!("Normalized shape: {:?}", normalized.shape());
    Ok(())
}

PCA

use scirs2_transform::reduction::PCA;
use scirs2_core::ndarray::Array2;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let data = Array2::<f64>::zeros((200, 50)); // high-dimensional data

    let mut pca = PCA::new(10, true, false); // 10 components, center=true
    let reduced = pca.fit_transform(&data)?;

    if let Some(evr) = pca.explained_variance_ratio() {
        let total: f64 = evr.iter().take(10).sum();
        println!("Explained variance (10 components): {:.1}%", total * 100.0);
    }
    println!("Reduced shape: {:?}", reduced.shape()); // (200, 10)
    Ok(())
}

UMAP

use scirs2_transform::umap::UMAP;
use scirs2_core::ndarray::Array2;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let data = Array2::<f64>::zeros((500, 100));

    let mut umap = UMAP::new(2)     // 2D embedding
        .with_n_neighbors(15)
        .with_min_dist(0.1);
    let embedding = umap.fit_transform(&data)?;

    println!("UMAP embedding shape: {:?}", embedding.shape()); // (500, 2)
    Ok(())
}

Persistent Homology (TDA)

use scirs2_transform::tda::{VietorisRips, PersistentHomology};
use scirs2_core::ndarray::Array2;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let points = Array2::<f64>::zeros((50, 3));

    let vr = VietorisRips::new(2.0, 1); // max_radius, max_dim
    let complex = vr.build(points.view())?;

    let ph = PersistentHomology::new();
    let diagrams = ph.compute(&complex)?;

    println!("H0 features: {}", diagrams[0].len());
    println!("H1 features: {}", diagrams[1].len());
    Ok(())
}

Optimal Transport

use scirs2_transform::optimal_transport::{sinkhorn, wasserstein_distance};
use scirs2_core::ndarray::{Array1, Array2};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let source = Array1::<f64>::from_vec(vec![0.5, 0.5]);
    let target = Array1::<f64>::from_vec(vec![0.3, 0.7]);
    let cost = Array2::<f64>::from_shape_vec((2, 2), vec![0.0, 1.0, 1.0, 0.0])?;

    // Sinkhorn regularized OT (reg=0.1)
    let (transport_plan, ot_cost) = sinkhorn(
        source.view(), target.view(), cost.view(), 0.1, 100
    )?;
    println!("OT cost: {:.4}", ot_cost);
    Ok(())
}

Feature Flags

Related Crates

• scirs2-cluster - Clustering algorithms
• scirs2-linalg - Linear algebra (SVD, eigendecomposition)
• scirs2-fft - FFT operations (used by signal transforms)
• SciRS2 project

License

Licensed under the Apache License, Version 2.0. See LICENSE for details.