For forgetting = 1, the processor retains Chan-Golub-LeVeque sufficient
statistics and is batch-equivalent without retaining samples. Smaller
forgetting factors produce an explicitly exponentially weighted covariance.
Usage
incrementalPCA(n_components, n_features = NULL, forgetting = 1, center = TRUE)
Arguments
- n_components
Number of retained components.
- n_features
Optional feature count to bind at construction.
- forgetting
Per-sample covariance forgetting factor in (0, 1].
- center
Whether to estimate and subtract a running mean.
Value
A mutable IncrementalPCA streaming processor.
Examples
set.seed(1)
latent <- matrix(rnorm(300), ncol = 3)
samples <- latent %*% matrix(rnorm(12), 3, 4)
pca <- incrementalPCA(n_components = 2L)
result <- update(pca, samples)
result$diagnostics$explained_variance
#> [1] 8.033294 4.948562