Score the self-gated mixture-of-experts single-sample discriminator

Scores each row of X independently with the frozen model from fit_moe_gated, in PURE R (no python). Each query is mapped to the per-sample robust CLR over the frozen feature universe (absent universe features carry the neutral rCLR value 0), embedded by the exported LayerNorm encoder forward to the pre-activation embedding $z$, and scored by the self-gated mixture logit $s(z) = \sum_e g_e(z)\,h_e(z)$. Larger = more case-like.

Queries with fewer than model$hp$min_features universe features present (a column-overlap floor), or empty positive support over the frozen universe on the (pre-rCLR) aligned abundances (!any(X_use[i, ] > 0)), return the neutral score 0. A FLAT all-equal-positive composition maps to the rCLR origin but is a VALID specimen and is scored normally. The score of a row depends only on that row and the frozen model and is exactly invariant to per-specimen positive scaling.

Usage

score_moe_gated(model, X, meta = NULL)

Arguments

model: A moe_gated_model object from fit_moe_gated.
X: Numeric matrix (samples $\times$ features) of non-negative abundances with named feature columns.
meta: Optional per-sample metadata. Accepted for interface uniformity and ignored by this method (the gate is self-driven).

Value

Plain finite numeric vector of length nrow(X); larger values are more case-like.

References

Jacobs RA, et al. (1991) Adaptive Mixtures of Local Experts. Neural Computation 3(1):79-87. Shazeer N, et al. (2017) Outrageously Large Neural Networks. ICLR. arXiv:1701.06538.

Examples

if (FALSE) { # \dontrun{
model <- fit_moe_gated(X, y)
score_moe_gated(model, X[1, , drop = FALSE])
} # }