The nearest prototype is pulled toward a training point when their labels
agree and pushed away when they do not. That repulsion is the whole
difference from k-means-per-class: it drives prototypes OUT of the contested
region, so where classes overlap the fitted prototypes end up further apart
than the class centroids. They are placed to win the nearest-prototype vote,
not to summarise their class.
Usage
morie_esl_prototype_lvq(
X,
y,
n_prototypes = 2,
eta = 0.1,
n_epochs = 50L,
newdata = NULL,
seed = 0L
)
Arguments
- X
Predictors, n by p.
- y
Class labels.
- n_prototypes
Prototypes per class.
- eta
Initial learning rate in (0, 1].
- n_epochs
Passes over the data.
- newdata
Points to classify; defaults to X.
- seed
Seed.
Value
List with prototypes, prototype_class, class, accuracy.
Details
LVQ is defined by an algorithm rather than an objective, so nothing is
guaranteed to decrease; the learning rate is decayed to zero to force the
iteration to settle.
References
Kohonen, T. (1989). Self-Organization and Associative Memory
(3rd ed.). Springer.
Examples
set.seed(1)
X <- rbind(matrix(rnorm(200, -2), ncol = 2), matrix(rnorm(200, 2), ncol = 2))
y <- rep(0:1, each = 100)
morie_esl_prototype_lvq(X, y, n_prototypes = 2, seed = 1L)$accuracy > 0.9
#> [1] TRUE
