analysis op• データ種: matrix × signal → table
• 呼び出し: import fullseye as fs; fs.ledger.mv_pls(X=None, Y=None, n_components=None, max_components=10, cv=7, method='nipals', standardize=True) -> 'dict' (実装を直接呼ぶなら import multivar; multivar.mv_pls(X=None, Y=None, n_components=None, max_components=10, cv=7, method='nipals', standardize=True) -> 'dict'、台帳から引くなら opsmultivar.get("mv_pls"))
Y を X に回帰する部分最小二乗(交差検証の Q^2、Q^2 による成分数の規則、VIP つき)。
> 以下の詳細説明は原文のままです —— 要約と見出しは訳出済み。
PLS finds directions in X that both vary and predict Y — the tool for many correlated predictors and few runs (image features -> process settings, or the reverse). method="nipals" (default, Wold, Sjöström & Eriksson 2001) or "simpls" (de Jong 1993; identical to NIPALS for a single response — a test pins it). X and Y are centred and, by default, scaled to unit variance.
Q^2 is computed by cv-fold cross-validation with interleaved folds (run i in fold i mod cv, as in SIMCA's default of 7 groups): Q^2_a = 1 - PRESS_a / SS_{a-1} per component (SS_0 = total sum of squares of Y, SS_{a-1} the residual sum of squares with a-1 components), cumulative Q^2 = 1 - PRESS_A / SS_0. n_components=None keeps adding components while Q^2_a >= 0.0975 (= 1 - 0.95^2; Wold et al. 2001) up to max_components; if not even the first passes, one component is fitted and significant is False — the data then do not support a predictive model, and the coefficients should not be used.
VIP (variable importance in projection) ranks X columns; the conventional threshold is VIP > 1. coef / intercept predict Y in its original units from X in its original units.
• design_of_experiments ファミリ ガイド
• サンプルデータ カタログ(DL URL / ライセンス) — 2-D は skimage.data(BSD/public)+ 合成、3-D は実データ源(Stanford/PDS 等)の DL URL。
• 演算子の来歴・参考文献 — この op 族の元になった研究/手法の出典。
• アルゴリズムの正典(著者・年)と用途は上記ファミリ使い方ガイドに記載。
• multivar_inspection_tour — py -3.11 examples/multivar_inspection_tour.py
table を入力に取れる)—
analysis)mv_pca · rsm_fit · doe_main_effects
*Provenance: multivar.py — MULTIVAR operator registry. この per-op ノートは tools/opdocs.py md が自動生成(手編集しない)。*
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.