metaflowx 1.1.4

A comprehensive, modular machine-learning framework that consolidates datasets, model architectures, optimization algorithms, preprocessing utilities, and evaluation tools into a unified toolkit designed for research, experimentation, and education. metaflowx streamlines end-to-end workflows—from data ingestion to model selection, training, tuning, and validation—empowering learners and researchers with a clean, extensible, and production-aware interface.

metaflowx

A modular, enterprise-grade machine-learning library engineered to streamline data workflows, accelerate model development, and operationalize end-to-end analytics pipelines. The framework is purpose-built to enable research-grade experimentation while maintaining production-level governance, reproducibility, and scale-out capability.

---

Overview

metaflowx is positioned as a full-stack machine-learning toolkit designed to deliver high operational efficiency across the entire data lifecycle. The package consolidates industry-standard modeling utilities, modern preprocessing pipelines, advanced decomposition operators, robust ensemble systems, and battle-tested evaluation modules.
The architecture is aligned for extensibility, maintainability, and high-performance execution, empowering both academic research and industrial ML deployments.

The solution leverages a structured module layout inspired by modern ML ecosystems, accelerating onboarding and cross-functional collaboration. Its broad coverage allows teams to build, benchmark, and operationalize ML assets with minimized technical debt and maximum throughput.

Note: This README intentionally excludes detailed documentation for the optimiser folder functions, as requested.

---

Key Value Proposition

• End-to-end workflow acceleration—curated building blocks covering datasets, feature engineering, model selection, and post-training evaluation.
• Scalable and modular architecture—clean separation of responsibilities, reusable primitives, and enterprise-friendly structure.
• High test coverage—extensive test suites ensure deterministic execution and strong governance across ML releases.
• Research-ready and production-aligned—balances academic flexibility with corporate-grade engineering rigor.
• Dataset-first philosophy—built-in access to canonical datasets enables rapid prototyping and standardized benchmarking.

---

Project Structure

Datasets Module (datasets/)

A comprehensive library of dataset loaders, parsers, readers, and test datasets.

Capabilities

• Native support for canonical datasets including Iris, Breast Cancer, Wine, California Housing, LFW, 20 Newsgroups, RCV1, KDDCup99, and Species Distributions.
• Local data ingestion layer with CSV/ARFF/SVMLight support.
• Internal OpenML readers to streamline API-free reproducibility for experimentation.
• Optimized SVMLight parsers underpinned by Cython for throughput and reliability.

Subcomponents

• data/: prepackaged datasets (CSV, GZ archives).
• descr/: structured documentation describing metadata and schema for each dataset.
• images/: bundled sample images for vision workflows.
• tests/: dataset-level validation ensuring schema consistency and deterministic outcomes.
• _svmlight_format_fast.pyx: Cython-accelerated parsing engine for large-scale ingestion.

The dataset hub enables fast onboarding, standardized evaluation cycles, and simulation of production-scale ingestion patterns.

---

Decomposition Module (decomposition/)

A full suite of dimensionality-reduction and matrix-factorization utilities engineered for speed and analytical clarity.

Operators Include:

• PCA, Incremental PCA, Kernel PCA
• FastICA
• Factor Analysis
• Sparse PCA
• Truncated SVD
• Non-Negative Matrix Factorization (NMF)
• Online LDA (Latent Dirichlet Allocation)
• Dictionary Learning
• Coordinate Descent NMF (Cython-accelerated)

Performance Considerations

• High-throughput Cython kernels (_cdnmf_fast.pyx, _online_lda_fast.pyx)
• Numerical stability enhancements for large datasets
• Memory-efficient incremental methods suitable for streaming workloads

Ideal for feature extraction, representation learning, signal separation, and topic modeling across structured and unstructured datasets.

---

Ensemble Module (ensemble/)

A complete ensemble-learning suite instrumented for operational robustness and model governance.

Supported Frameworks:

• Random Forests
• Extra Trees
• Bagging
• Boosting (Gradient Boosting, Histogram-Based Gradient Boosting)
• Stacking
• Voting
• Isolation Forest for anomaly detection

Engineering Highlights:

• Cython-optimized histogram-based boosting stack (_hist_gradient_boosting/)
• Bitset-based performance enhancements for categorical splits
• Monotonic constraint support for regulated industries (finance, healthcare)
• Predictive pipeline fully aligned with high-volume production workloads

The ensemble stack ensures consistent, scalable model performance even in high-cardinality environments.

---

Feature Selection Module (feature_selection/)

A research-grade toolkit for filter, wrapper, and embedded feature-selection strategies.

Tooling Includes:

• Variance Thresholding
• Univariate Statistical Tests
• Mutual Information
• RFE / RFECV
• Sequential Feature Selectors
• Model-based selectors
• Cython-driven mutual information estimators

Built for teams optimizing feature pipelines, reducing computational overhead, and enhancing interpretability.

---

Frozen Models Module (frozen/)

A controlled environment for immutable model artefacts used in audit-compliant ML pipelines.

This module supports freezing model states to enforce reproducibility during validation or regulatory review.

---

Linear Models (linear_model/)

An industrial-strength suite of regression and classification algorithms.

Coverage:

• Ridge, Lasso, ElasticNet
• Logistic Regression
• Bayesian Regression
• SGD-based solvers
• Huber Regression
• Quantile Regression
• Coordinate Descent engines
• Passive-Aggressive models
• Least Angle Regression (LARS)
• RANSAC Robust Regression
• Theil–Sen Estimator

Cython integrations (_cd_fast.pyx, _sag_fast.pyx.tp, _sgd_fast.pyx.tp) provide scale-up capability for enterprise workloads.

---

Model Selection (model_selection/)

A holistic module for model tuning, split strategies, and performance validation.

Functional Areas:

• Train/Validation/Test split orchestration
• K-Fold, Stratified K-Fold, Shuffle Splits
• Grid Search, Random Search, Successive Halving
• Classification threshold optimization
• Visualization utilities for diagnostic analysis (_plot.py)
• Enhanced validation logic with deterministic behaviors

The module is designed for repeatable experimentation and strong audit trails.

---

Neural Network Module (neural_network/)

A lightweight neural network stack built around core feedforward architectures.

Included:

• MLP classifiers and regressors
• RBM (Restricted Boltzmann Machine)
• Stochastic optimization utilities
• Gradient-based solvers tuned for controlled training regimes

Positioned as a research accelerator rather than a deep learning framework.

---

Preprocessing (preprocessing/)

A scalable data-preprocessing library that minimizes friction in ETL and feature engineering pipelines.

Assets Include:

• Label Encoding, One-Hot Encoding, Ordinal Encoding
• Target Encoding (with Cython-accelerated fast path)
• Polynomial Feature Expansion
• Binning and Discretization
• Scalable sparse matrix transformations
• Function transformers for custom data logic

This module anchors the data engineering pipeline, enabling clean, repeatable transformations.

---

Support Vector Machines (svm/)

A feature-rich SVM implementation built on top of optimized C++ backends.

• LibSVM and LibLinear integrations
• Sparse SVM routines
• Cython bridges for accelerated inference
• Deterministic linear SVM solvers
• C++ template architecture for compute-efficient training

This subsystem enables enterprise teams to deploy classical ML with predictable performance and full reproducibility.

---

Tests

The repository includes an extensive automated testing framework across all modules, ensuring:

• Regression safety
• Deterministic output behavior
• Conformance to expected data and model interfaces
• Compliance with enterprise release processes

CI-friendly test structure enables frictionless integration into DevOps pipelines.

---

Installation

Standard installation via pip:

pip install metaflowx