bigPLSR (version 0.7.2)

kf_pls_state_new: KF-PLS streaming state (constructor)

Description

Create a persistent Kalman–filter PLS (KF-PLS) state that accumulates cross-products from streaming mini-batches and later produces a big_plsr-compatible fit via kf_pls_state_fit().

Usage

kf_pls_state_new(p, m, ncomp, lambda = 0.99, q_proc = 0, r_meas = 0)

Value

An external pointer to an internal KF-PLS state (opaque object) that you pass to kf_pls_state_update() and then to kf_pls_state_fit() to produce model coefficients.

Arguments

p

Integer, number of predictors (columns of X).

m

Integer, number of responses (columns of Y).

ncomp

Integer, number of latent components to extract at fit time.

lambda

Numeric in (0,1], forgetting factor (closer to 1 = slower decay).

q_proc

Non-negative numeric, process-noise magnitude (adds a ridge to \(C_{xx}\) each update; useful for stabilizing ill-conditioned problems).

r_meas

Reserved measurement-noise parameter (not used by the minimal API yet; kept for forward compatibility).

Details

The state maintains exponentially weighted cross-moments \(C_{xx}\) and \(C_{xy}\) with forgetting factor lambda. When lambda >= 0.999999 and q_proc == 0, the backend switches to an exact accumulation mode that matches concatenating all chunks (no decay).

See Also

kf_pls_state_update(), kf_pls_state_fit(), pls_fit() (use algorithm = "kf_pls" for the one-shot dense path).

Examples

set.seed(1)
n <- 1000; p <- 50; m <- 2
X1 <- matrix(rnorm(n/2 * p), n/2, p)
X2 <- matrix(rnorm(n/2 * p), n/2, p)
B  <- matrix(rnorm(p*m), p, m)
Y1 <- scale(X1, TRUE, FALSE) %*% B + 0.05*matrix(rnorm(n/2*m), n/2, m)
Y2 <- scale(X2, TRUE, FALSE) %*% B + 0.05*matrix(rnorm(n/2*m), n/2, m)

st <- kf_pls_state_new(p, m, ncomp = 4, lambda = 0.99, q_proc = 1e-6)
kf_pls_state_update(st, X1, Y1)
kf_pls_state_update(st, X2, Y2)
fit <- kf_pls_state_fit(st)          # returns a big_plsr-compatible list
preds <- predict(bigPLSR::.finalize_pls_fit(fit, "kf_pls"), rbind(X1, X2))
head(preds)