Figure 4.15. Multiway multiblock regression problem [297, 556].
variance in Y accounted for by T: a ■ R2XT + (1 — a) • RyT, with 0 < a < 1. The least-squares loss function may be written in terms of the residuals as a minimization problem:
o-(W,Px,Py) =a||X-XWPx||2 + (l-a)||Y-XWPy||2 (4.156)
with the Frobenius matrix norm || • || and constraint TtT = WTXTXW = 1a, where is an A x A identity matrix. The method can be extended to multi-block data. While de Jong and Kiers  consider prediction of Yfc, k > 1 from a single X, it is also possible to formulate problems where many X blocks are used to predict Y.
Was this article helpful?