Module learning 

Statistical learning (ISL) — predictive estimators built on the engine’s existing foundation, never duplicating it (see stats_plan.md).

Reuses: solvers::linear_algebra (gemm/qr/cholesky/eigen/svd) for the linear algebra, solvers::statistics (descriptive/distributions/correlation) for moments and p-values, and platform::compute_bridge for per-kernel-class dispatch.

Categories (one method-family per sub-library, PROJECT RULE §13): metrics, preprocessing, regression, glm, classification, resampling, dimensionality, clustering, trees, splines, survival, multiple_testing.

Modules

active

Active learning — spend the human’s attestation budget wisely.

classification

Classification (ISL ch 4) — generative and instance-based classifiers.

clustering

Clustering & mixture models (ISL ch 12, PRML ch 9).

dimensionality

Dimensionality reduction (ISL ch 12, PRML ch 12). Built on linear_algebra (eigen/SVD). Feeds the engine’s 10D→5D NQuin relevance router.

experiment

Experiment design & evaluation (Practical Statistics ch 3) — planning and analysing experiments, with uncertainty made explicit.

gaussian_process

Gaussian Process regression (PRML ch 6.4) — a nonparametric Bayesian regressor that returns a full predictive distribution (mean, variance) at every input, with the squared-exponential (RBF) kernel. The training solve reuses linear_algebra::cholesky (no new solver). Kernel-class DenseLinear (the n×n kernel solve) + AllPairs (the kernel evaluations).

glm

Generalized linear models (ISL ch 4) — logistic and Poisson regression by iteratively reweighted least squares (IRLS).

graphical_models

Probabilistic graphical models (PRML ch 8) — relational inference over discrete variables. The governance topology is relational and lives in the edges, so a factor graph with belief propagation is the native inference for it.

kg_embedding

Knowledge-graph embedding (TransE / DistMult / ComplEx / RotatE) — score a triple (head, relation, tail) for plausibility, and rank candidate entities for link prediction over the semantic graph.

metrics

Model-evaluation metrics for statistical learning (ISL ch 2, 4, 9).

multiple_testing

Multiple-testing corrections (ISL ch 13) — adjust a set of p-values for the number of hypotheses tested, controlling either the family-wise error rate (Bonferroni, Holm) or the false discovery rate (Benjamini–Hochberg).

preprocessing

Preprocessing for statistical learning — feature scaling and design-matrix prep. Reuses statistics::descriptive. Train/test partitioning lives with the resampling module (it shares the fold/shuffle machinery).

regression

Regression estimators (ISL ch 3, 6). Built on linear_algebra + statistics.

resampling

Resampling methods (ISL ch 5) — cross-validation and the bootstrap. The generic harness every later chapter reuses to estimate test error / variability without a separate validation set.

sampling

Monte-Carlo sampling (PRML ch 11) — the inference engine for Bayesian methods.

sequential

Sequential / temporal models (PRML ch 13) — estimators over time-indexed data.

splines

Regression splines & polynomial regression (ISL ch 7) — flexible non-linear fits expressed as a linear model in a fixed basis, then solved by OLS (learning::regression::linear, no new solver).

survival

Survival analysis (ISL ch 11) — time-to-event modelling with right censoring. Time-indexed provenance / life-record reasoning is temporal, so these are the standard estimators over censored temporal evidence.

trees

Tree-based methods (ISL ch 8) — CART trees and their ensembles.

variational

Variational inference (PRML ch 10) — approximate an intractable posterior by the closest factorized distribution (mean-field), via coordinate-ascent (CAVI).

Enums

LearningError

Errors common to the learning estimators. Estimators fail closed (return an error) rather than emit a fabricated fit — consistent with the engine-wide honesty rule.