[LaurianeJD]

Purpose

Fix NipalsPLS.predict() when y has missing values, as discussed in #14 (see comment there for more details).

Background

The fit deflates y with T Qᵀ, but predict() returned T diag(b) Qᵀ. The inner coefficients b are exactly 1 for complete y, but not when y has missing values, where predict() then doesn't reproduce the fitted y, even on the training data.

Changes

• predict(): Ŷ = T Qᵀ, without regression_matrix (b term)
• get_reg_vector(): same, without regression_matrix
• New test test_predict_with_missing_y(): fails on main, passes with the fix

The test uses a small random example rather than existing test data: YData.csv has a single complete y column (b = 1). A second target built from the spectra with missing values (as in #14) shows the problem only weakly (b ≈ 1.0004), while this random example shows it clearly (b = 3.4).

Results with complete y are unchanged.

Caveat

Overlaps with #14: both PRs change get_reg_vector(), and this fix also covers #14's set_components() growth fix with missing y. See my comment there.

@nielsschlusser @arden13

[nielsschlusser]

Looks good to me in general!
One minor question with regards to the test, otherwise good to merge from my side.
Let's wait for PR #14 , though.

[nielsschlusser]

Looks good to me now. Thanks for your contribution!
Please wait with merging until the merge of PR #14 .

[LaurianeJD]

@nielsschlusser Sure! After PR #14 is merged, I'll also add here the remaining necessary changes that I listed in #14 (comment) to make sure everything is consistent :)

[nielsschlusser]

Hey @LaurianeJD

I just had another thought: Shall we try to tighten the limits for the prediction unittest? I'd imagine the accuracy on the incomplete datasets should be increased. What do you think?

[LaurianeJD]

@nielsschlusser As far as I understood, test_fit_y isn't affected by my PR: the incomplete datasets used in that test have missing values in the features X, not in the target y. Since for that case y is complete, b=1 exactly, and the predictions are identical with and without the fix of this PR.
Or am I misunderstanding what you were mentioning?