.. _auto_dataset_pca_fit:

*DATASET PCA (FIT)*
===================

Fits PCA on a set of numeric dataset columns -- selected by a column name prefix (e.g. 'MEAS_ENGSS_YY_'), an explicit list, or BOTH combined (the union) so columns from different patterns can be decomposed together. Appends the principal-component scores as new columns ('<pc_prefix>_PC1'..'_PCn') and emits the fitted projection (components/mean/std/columns) as a reusable parameters object. Standardizes over available values only; missing values handled by the chosen strategy ('zero' = fill with training mean; 'iterative' = iterative SVD imputation). SVD via numpy. Use on the TRAINING dataset, then feed 'pca_parameters' into dataset_pca_transform and dataset_pca_inverse.

When to use
-----------

Tagged: ``pca``, ``dimensionality_reduction``, ``svd``, ``fit``, ``decomposition``, ``preprocessing``, ``curves``, ``iterative_svd``.

Inputs
------

.. list-table::
   :header-rows: 1
   :widths: 20 20 20 20 20 20

   * - Label
     - ID
     - Type
     - Default
     - Required
     - Description
   * - Dataset
     - dataset
     - dataset
     - —
     - ✓
     - Training dataset with the curve columns to decompose; array-of-row-objects.
   * - Column Prefix
     - column_prefix
     - text
     - —
     - 
     - Prefix selecting numeric columns to decompose, e.g. 'MEAS_ENGSS_YY_'. Columns whose remainder after the prefix is an integer are used, sorted numerically. Can be combined WITH the explicit Columns list (the union is decomposed).
   * - Columns
     - columns
     - text
     - —
     - 
     - Explicit list of additional columns to decompose, e.g. MEAS_ENGSS_xfirst, MEAS_ENGSS_xlast. Combined with Column Prefix (union) -- use this to add columns that don't share the prefix pattern.
   * - PC Column Prefix
     - pc_prefix
     - text
     - PC
     - 
     - Prefix for the emitted score columns, e.g. 'EngSS' -> EngSS_PC1, EngSS_PC2, ... Must match what dataset_pca_inverse expects.
   * - Number of Components
     - n_components
     - number
     - 10
     - 
     - Number of principal components to keep (clamped to the achievable rank).
   * - Missing Value Strategy
     - missing_strategy
     - select
     - zero
     - 
     - How to handle missing values when fitting. 'zero': missing standardized values filled with 0 (the training mean), single SVD. 'iterative': iterative SVD imputation (re-impute missing cells from a low-rank estimate, repeat) with per-row least-squares scores -- better when columns are heavily missing (e.g. curves with shorter x-ranges).

Outputs
-------

.. list-table::
   :header-rows: 1
   :widths: 20 20 20 20

   * - Label
     - ID
     - Type
     - Description
   * - Dataset With PCs
     - dataset
     - dataset
     - Input dataset with appended <pc_prefix>_PC1..n score columns.
   * - PCA Parameters
     - pca_parameters
     - json
     - Fitted projection: columns, mean, std, components, pc_prefix, missing_strategy, explained_variance_ratio. Reuse with dataset_pca_transform / dataset_pca_inverse.
   * - Status
     - status
     - string
     - Summary of columns decomposed and components kept.

Disciplines
-----------

- data.dataset.transform
- data.statistics

.. raw:: html

   <hr style="margin-top:2em">
   <p style="font-size:11px;color:#888">
   Auto-generated from <code>platform</code> schema. Worker id: <code>dataset_pca_fit</code>. Schema hash: <code>7c83511128c3</code>. Hand-curated docs in <code>workerexamples/</code> override this page when present.
   </p>