tabnetics 1.0.0

HDLSS-focused tabular learning toolkit with distribution-aware preprocessing, portfolio feature selection, and game-theoretic method aggregation.

Tabnetics

A Python toolkit for high-dimensional, low-sample-size (HDLSS) tabular classification. Tabnetics grew out of the review paper Machine learning on small size samples: A synthetic knowledge synthesis, which provided the library's initial theoretical background for small-sample learning. The library combines distribution-aware preprocessing, portfolio-based feature selection, and game-theoretic method aggregation into a single pipeline designed for settings where p >> n.

Homepage: tabnetics.org

pip install tabnetics

This base install now matches the ready-to-run public runtime surface and ships every currently packaged direct dependency except TabPFN. For the fully loaded opt-in stack, including TabPFN, use pip install "tabnetics[full]".

Licensed under Apache 2.0.

Tabnetics depends on several third-party libraries with separate licenses and use terms. Those upstream terms still apply when you install or enable the related Tabnetics feature. Review Third-party integrations and licenses before using the packaged dependencies or the full extra.

What's new in 1.0.0

First stable release. Tabnetics moves from beta to production-stable with version 1.0.0, reflecting four completed validation campaigns (Val-18 through Val-21) and a mature pipeline surface. The public benchmark evidence base now covers 56,107 successful runs across 278 pipeline profiles on 63 HDLSS datasets. The best profiles clear 36/63 datasets at balanced accuracy >= 0.90, achieve 26 perfect-classification datasets, and land 41 / 21 / 1 in the strict-holdout SOTA split (above / within / below).

Val-21 integrated. The newest validation campaign (Val-21) adds 7 decision-focused profiles testing winner-bridge, meta-selector, prefilter, and regime variants across 43 datasets. Val-21 results are now included in the interactive results browser and the aggregate evidence base alongside Val-18/Val-19/Val-20.

Cross-campaign technical report. A unified technical report covering the full Val-18/Val-19/Val-20 evidence base is available. The report presents matched-overlap paired comparisons with Wilcoxon signed-rank tests, per-tier decomposition, and a pre-registered hypothesis scorecard covering 22 testable hypotheses across the three campaigns. Key findings: BH filtering is strongly validated as a safety feature (disabling costs −0.046 BA on 63 datasets, p < 1e-4), FLAML at 60 s is the cleanest frontier improvement (+0.003 BA, p = 0.017), and the full FS+DF pipeline scaffold remains directionally beneficial.

Expanded classifier surface. The HDLSS extreme-regime pool includes 28 classifiers and the moderate-regime pool includes 29, covering bias-corrected DA/SVM variants, CPDA, copula-style and robust classifiers, RFF / nearest-subspace geometries, HDRDA-style regularized discriminants, DWD, sparse PLS-DA, ECOC lifts, and the optional full-fidelity pytabkit paths for TabM and RealMLP.

Import-safety fix. Module-level find_repo_root() calls in the validation subsystem no longer crash when the package is installed via pip outside a source checkout.

When to use Tabnetics

Tabnetics is built for tabular classification problems where the number of features greatly exceeds the number of samples:

• Transcriptomics — microarray and RNA-seq gene expression
• Proteomics and metabolomics — mass-spec feature matrices
• Other HDLSS settings — any structured tabular problem with p >> n

In these regimes the dominant failure modes are not model selection — they are unstable preprocessing, brittle feature selection, information leakage, and inflated validation estimates. Tabnetics addresses all four.

What Tabnetics adds to the HDLSS problem is not just another selector: it turns many unstable HDLSS choices into a multiplayer portfolio game. Feature-selection methods and classifier candidates are treated as competing players, oracle scores become the payoff structure, and the resulting MNPO equilibrium is used to select a robust portfolio under small-sample constraints.

Usage guide → · Methods & references → · Benchmark results → · Results browser → · Announcements

Call for collaboration

We are actively looking for testers, collaborators, and co-authors to help validate Tabnetics on real-world HDLSS datasets, shape the companion article, and improve the codebase. If you work with high-dimensional tabular data — transcriptomics, proteomics, metabolomics, or similar — we would love to hear from you. See the Discussions page for ongoing conversations, or open a new thread to introduce your use case.

Citation

If you use Tabnetics in research, cite the repository for the specific version you used. The library is still under active development, and a companion paper will be published after the current testing and validation cycle is complete.

Repository URL: https://github.com/klokedm/tabnetics-public

@software{kokol_tabnetics_2026,
  author = {Kokol, Marko},
  title = {Tabnetics},
  year = {2026},
  url = {https://github.com/klokedm/tabnetics-public}
}

Benchmark results

Tabnetics has been evaluated on 63 HDLSS benchmark datasets (41–7,000 samples, 500–100,001 features, 2–14 classes) drawn from OpenML, GEO, CuMiDa, Scikit-feature, and UCSC Xena/TCGA. Across the consolidated Val-18 through Val-21 evidence base, the public results now summarize 56,107 runs with 278 pipeline profiles. The best profiles exceed published strict-holdout SOTA on 41 of 63 datasets (65%) and match or exceed published ranges on 62 of 63 (98%). Twenty-six datasets achieve perfect classification and 36 exceed 0.90 balanced accuracy. Detailed per-dataset results, statistical comparisons, and article references are available in RESULTS.md, and the static results browser exposes the same public surfaces as interactive charts and tables. A cross-campaign technical report with full methodology and ablation studies is available, and a peer-reviewed article is in preparation.

Key ideas

1. Distribution-aware preprocessing. Each feature is fitted to a parametric family (from 20+ candidates) using goodness-of-fit testing, bootstrap calibration, and L-moment prescreening. CDF-based transforms replace ad-hoc normalization.

2. Portfolio feature selection. Forty feature-selection paths are available overall, including 39 engineered selectors plus a random-baseline reference used in validation. The benchmark MNPO portfolios combine stability selectors, copula knockoffs, tree-based importance, mutual-information filters, IPSS, HSIC-Lasso, and more into a single robust HDLSS feature portfolio. MNPO builds pairwise preference matrices from multiple oracles (performance, stability, complexity, etc.) and solves for a Nash equilibrium via KL-regularized mirror descent. The multiplayer game framing draws conceptual inspiration from Wu et al.'s Multiplayer Nash Preference Optimization, though the HDLSS adaptation is a distinct contribution with different players, oracles, and data regime (see BACKGROUND.md for details).

3. Regime-aware classification. An MNPO-based classifier oracle picks from regime-appropriate pools. The HDLSS extreme pool now spans 28 candidate backends, including linear/GLM baselines, LDA/QDA families, bias-corrected HDLSS discriminants, PLS and sparse PLS variants, random-projection and random-Fourier lifts, nearest-subspace and spatial-median geometry, copula-style discriminants, confusion-pursuit, DWD / ECOC wrappers, and lightweight or full-fidelity deep-tabular models (TabM, RealMLP). The HDLSS moderate pool carries 29 backends by adding RBF SVM, GPC, KNN, vote ensembles, and TabPFN on top of the HDLSS-safe core. Standard-regime routing can also unlock tree families such as RF, Extra Trees, XGBoost, LightGBM, and CatBoost. The base package now includes the always-shipped boosted-tree and pytabkit backends directly; TabPFN remains the only kept opt-in classifier path, and custom environments still degrade gracefully if any third-party backend is absent.

4. Strict validation. All learned preprocessing and selection is train-only. The HuggingFace bundle is the authoritative reproducibility mirror of the public upstream datasets used for validation. Synthetic fallback is not allowed for evidence-bearing runs.

Quick start

The quick-start path assumes the recommended expanded install above, so the broader optional selector and backend surface is already available when you need it.

from tabnetics.pipeline import DistributionFeatureSelectionPipeline, DFFSConfig

config = DFFSConfig(random_seed=42)
pipeline = DistributionFeatureSelectionPipeline(config)

result = pipeline.run(X, y, dataset_name="my_dataset", seed=42)

print(f"Accuracy: {result.accuracy:.3f}")
print(f"Selected features: {result.selected_features}")

Operational defaults

The packaged runtime currently follows the promoted post-review workflow:

• df_stage_position="after_fs" is the default, so distribution fitting runs on the feature space that actually survives selection.
• Evidence-bearing benchmark and validation runs treat the HuggingFace bundle as the authoritative reproducibility mirror of the public upstream datasets and default to dataset_integrity_policy="error".
• Conformal prediction is opt-in and should be interpreted as an uncertainty layer (coverage, prediction-set size, singleton rate), not as a balanced-accuracy optimizer.
• multiomics_adapter="split_halves" is a benchmark-time shortcut; real multi-omics studies should use explicit blocks with tabnetics.multiomics.

Command line workflows

Editable installs expose installed wrappers, and every wrapper has the same packaged python -m ... equivalent:

tabnetics-benchmark --datasets leukemia_golub --seeds 11 23 37
tabnetics-validation-plan --plan-kind validation17 --num-pods 4
tabnetics-validation-suite --dataset-sets fs_easy --seeds 11 23 37

The corresponding module entrypoints are:

• python -m tabnetics.benchmarks.cli
• python -m tabnetics.validation.generate_plan
• python -m tabnetics.validation.core.shard_runner
• python -m tabnetics.validation.suite

Selected literature anchors

The full methods table lives in BACKGROUND.md. For a quick orientation, these are the main papers behind the current public positioning:

• Kokol. Machine learning on small size samples: A synthetic knowledge synthesis — the original HDLSS review context behind the library.
• Freund & Schapire. Adaptive game playing using multiplicative weights — the mirror-descent / multiplicative-weights foundation used by the MNPO solver.
• Wu et al. Multiplayer Nash Preference Optimization — conceptual inspiration for the multiplayer Nash framing, not the solver implementation.
• Marron et al. Distance-Weighted Discrimination, Lê Cao et al. Sparse PLS discriminant analysis, and Dietterich & Bakiri Error-Correcting Output Codes — key anchors behind the expanded HDLSS classifier pool.
• Rahimi & Recht Random Features for Large-Scale Kernel Machines, Hall et al. Median-Based Classifiers for High-Dimensional Data, Tsuda Subspace classifier in the Hilbert space, and Han et al. CODA — the kernel, robust, subspace, and copula-style classifier families reflected in the current backend surface.
• Singh et al. DIABLO and Rohart et al. MINT — the reference points for explicit multi-omics integration.
• Taquet et al. MAPIE: an open-source library for distribution-free uncertainty quantification — the conformal/UQ reference behind the classifier-side uncertainty outputs.

Package structure

Subpackage	Purpose
tabnetics.core	MNPO game-theoretic primitives, sklearn compatibility layer, runtime configuration
tabnetics.distribution	Univariate distribution fitting (20+ families), bootstrap GOF, CDF-based transforms
tabnetics.feature_selection	30 selection methods, MNPO portfolio aggregation, copula knockoffs, stability selectors
tabnetics.classification	Regime-aware classifier pools, classical and specialist HDLSS backends, MNPO classifier oracle, and conformal helpers
tabnetics.pipeline	End-to-end DF+FS+classification pipeline with leakage prevention
tabnetics.datasets	Dataset registry, HuggingFace/OpenML loaders, meta-feature extraction
tabnetics.domains	Domain adapters (bioinformatics prefilters, face-domain projection)
tabnetics.multiomics	Multi-block PLS-DA (DIABLO-style) and MINT batch-correction integration
tabnetics.benchmarks	Benchmark runner, method-set profiles, SOTA comparison, gaming detection
tabnetics.validation	Validation campaign planner, shard execution, promotion gates

Feature selection methods

The FeatureSelector supports 40 methods out of the box, including 39 engineered selectors plus a random-baseline reference path:

Category	Methods
Stability selectors	Lasso stability, subspace stability, decorrelated stability, cluster stability, TIGRESS
Wrapper methods	RFECV (SVM, RF, LR), Boruta
Filter methods	ANOVA F-test, mutual information, mRMR, JMI, CMIM, FCBF, Wilcoxon AUC
Tree-based	GBDT importance, TreeSHAP, random forest
Knockoff methods	Copula knockoff (D-vine, FDR-controlled via e-values), derandomized knockoffs
Embedded	OA-Elastic Net, Joint AUC+L1, HSIC-Lasso
Other	IPSS, k-TSP, OVA/ECOC wrappers, Rashomon importance

Methods are aggregated via MNPO with configurable oracle presets (minimal, perf_only, perf_complexity, full, etc.).

See BACKGROUND.md for the full list of implemented papers, USING.md for detailed usage, and RESULTS.md for benchmark results.

Installation

Recommended install for the quick-start and usage-guide examples:

pip install tabnetics

Full install profile (same package plus the opt-in TabPFN path):

pip install "tabnetics[full]"

Requirements

• Python >= 3.11
• Plain tabnetics is the ready-to-run public runtime surface
• Optional install profile: full adds the TabPFN path

Development

git clone https://github.com/klokedm/tabnetics-public.git
cd tabnetics-public
pip install -e ".[dev]"
pytest

License

Apache 2.0 — see LICENSE.

---

This documentation is auto-generated from internal notes and sources with the support of rule-based transformations and generative AI. Errors are possible — please report any issues via Discussions.