- All Implemented Interfaces:
@Priority(-101) @Reference(authors="R. C. Jancey", title="Multidimensional group analysis", booktitle="Australian Journal of Botany 14(1)", url="https://doi.org/10.1071/BT9660127", bibkey="doi:10.1071/BT9660127") public class RandomNormalGenerated extends AbstractKMeansInitializationInitialize k-means by generating random vectors (normal distributed with \(N(\mu,\sigma)\) in each dimension).
This is a different interpretation of the work of Jancey, who wrote little more details but "introduced into known but arbitrary positions"; but seemingly worked with standardized scores. In contrast to
RandomUniformGenerated(which uses a uniform on the entire value range), this class uses a normal distribution based on the estimated parameters. The resulting means should be more central, and thus a bit less likely to become empty (at least if you assume there is no correlation amongst attributes... it is still not competitive with better methods).
Warning: this still tends to produce empty clusters in many situations, and is one of the least effective initialization strategies, not recommended for use.
R. C. Jancey
Multidimensional group analysis
Australian Journal of Botany 14(1)
- Erich Schubert
Nested Class Summary
Nested Classes Modifier and Type Class Description
All Methods Instance Methods Concrete Methods Modifier and Type Method Description
chooseInitialMeans(Relation<? extends NumberVector> relation, int k, NumberVectorDistance<?> distance)Choose initial means
Methods inherited from class elki.clustering.kmeans.initialization.AbstractKMeansInitialization
public RandomNormalGenerated(RandomFactory rnd)Constructor.
rnd- Random generator.
public double chooseInitialMeans(Relation<? extends NumberVector> relation, int k, NumberVectorDistance<?> distance)Description copied from interface:
KMeansInitializationChoose initial means
k- Parameter k
distance- Distance function
- List of chosen means for k-means