Tabnetics is organized around a leakage-safe HDLSS pipeline rather than a loose collection of selectors.

End-to-end flow

  1. tabnetics.datasets loads benchmark or user-supplied tabular data, keeps dataset metadata close to the run, and exposes catalog/meta-feature helpers used by both benchmarking and validation.
  2. tabnetics.distribution evaluates univariate distribution families, attaches fit diagnostics, and produces CDF-based transforms for the feature matrix. In the packaged runtime this stage defaults to df_stage_position="after_fs", so it operates on the post-selection feature space.
  3. tabnetics.feature_selection runs the configured selector portfolio, computes oracle signals, and aggregates selectors into an MNPO-weighted feature set.
  4. tabnetics.classification applies regime-aware classifier families and the classifier oracle to pick the downstream model for the selected features. Optional conformal diagnostics augment the classifier output with coverage and prediction-set information; they should be read as uncertainty metrics, not as point-accuracy optimizers.
  5. tabnetics.pipeline orchestrates the train/test split, audit rules, feature transformations, selection stages, classification, and reproducibility payloads.

Supporting modules

  • tabnetics.core holds shared runtime helpers, package-path discovery, and lower-level MNPO primitives that other packages build on.
  • tabnetics.multiomics provides explicit MINT and multi-block PLS helpers when the data is split into real omics blocks rather than a benchmark-style synthetic split; the benchmark-only split_halves adapter is just a convenience stress-test path.
  • tabnetics.domains keeps domain-specific routing and adapters separate from the generic pipeline surface.

Benchmarking and validation

  • tabnetics.benchmarks contains method-set profiles, CLI entrypoints, and the benchmark runner used for systematic comparisons.
  • tabnetics.validation builds campaign plans, shards work, runs validation suites, and records promotion-oriented evidence across benchmark catalogs.
  • Installed wrappers mirror the module entrypoints: tabnetics-benchmark, tabnetics-validation-plan, tabnetics-validation-shard, and tabnetics-validation-suite.
  • Evidence-bearing validation-catalog runs use the HuggingFace bundle as the authoritative operational mirror of the public upstream sources and default to dataset_integrity_policy=error; synthetic fallback is not part of the public validation path.

Practical reading order

Start with Using Tabnetics if you want to run the library, then read Methods and References for the algorithmic background, and finally use the Reference pages to inspect the package boundaries and source-linked API symbols.

Documentation and webpages on this site are generated from authoritative internal sources using a combination of deterministic rules and generative AI. Errors are possible. Please report issues via GitHub Discussions or email [email protected].