A=KVQ(H) returns a kvq object initialized with hyperparameters H.
The kvq object computes a learning vector quantization with guaranteed distortion
bounds. It does so by minimizing the L1 norm (parameter optmethod='l1lvq' / c.f. cited
reference below) or an approximation of the L0 norm (parameter optmethod='l0arom').
It also implements two variants of LVQ for labeled data: discriminative and shared
LVQ (set parameter mode to 'discriminative' or 'shared' respectively). In
discriminative LVQ a vector can only become a codebook vector if it has at least
distance delta (parameter distd) to any examples of the other class. In mode shared an
example can only become a codebook vector if it has at most distance delta to a example
of the opposite class.
Hyperparameters, and their defaults
dist=1; -- allowed point to point distortion.
distd=3 -- discriminative or shared radius
child=distance -- child stores the distance object.
a.cutoff = 1./3. -- cutoff value for importance values
a.return_indices = 0 -- return indices of points instead of samples
a.tol = 1e-6 -- tolerance parameter for matlab's LINPROG linear program optimizer
a.test_on_trainingset = 1 -- test the algorithm on the training data
Model
a.keep -- kept data points -- store data points kept for model
a.alpha -- importance factors
Methods:
train, test,circplot
Example:
d=gen(toy('n=2'));
[r,a]=train(kvq,d)
plot(a,d)
| |
Reference : A kernel approach vor vector quantization with guaranteed distortion bounds |
| |
Author : M. Tipping, B. Schölkopf |