elkans

Documentation

Index

• func L2Distance(v1, v2 *mat.VecDense) float64
• func NewKMeans(vectors [][]float64, clusterCnt, maxIterations int, deltaThreshold float64, ...) (kmeans.Clusterer, error)
• type ElkanClusterer
  • func (km *ElkanClusterer) Cluster() ([][]float64, error)
  • func (km *ElkanClusterer) InitCentroids() error
  • func (km *ElkanClusterer) SSE() float64
• type Initializer
  • func NewKMeansPlusPlusInitializer(distFn kmeans.DistanceFunction) Initializer
  • func NewRandomInitializer() Initializer
• type KMeansPlusPlus
  • func (kpp *KMeansPlusPlus) InitCentroids(vectors []*mat.VecDense, k int) (centroids []*mat.VecDense)
• type Random
  • func (r *Random) InitCentroids(vectors []*mat.VecDense, k int) (centroids []*mat.VecDense)

Functions

func L2Distance

func L2Distance(v1, v2 *mat.VecDense) float64

L2Distance is used for L2Distance distance in Euclidean Kmeans.

func NewKMeans

func NewKMeans(vectors [][]float64, clusterCnt,
	maxIterations int, deltaThreshold float64,
	distanceType kmeans.DistanceType, initType kmeans.InitType,
	normalize bool,
) (kmeans.Clusterer, error)

Types

type ElkanClusterer

type ElkanClusterer struct {
	// contains filtered or unexported fields
}

ElkanClusterer is an improved kmeans algorithm which using the triangle inequality to reduce the number of distance calculations. As quoted from the paper: "The main contribution of this paper is an optimized version of the standard k-means method, with which the number of distance computations is in practice closer to `n` than to `nke`, where n is the number of vectors, k is the number of centroids, and e is the number of iterations needed until convergence."

However, during each iteration of the algorithm, the lower bounds l(x, c) are updated for all points x and centers c. These updates take O(nk) time, so the complexity of the algorithm remains at least O(nke), even though the number of distance calculations is roughly O(n) only. NOTE that, distance calculation is very expensive for higher dimension vectors.

Ref Paper: https://cdn.aaai.org/ICML/2003/ICML03-022.pdf

func (*ElkanClusterer) Cluster

func (km *ElkanClusterer) Cluster() ([][]float64, error)

Cluster returns the final centroids and the error if any.

func (*ElkanClusterer) InitCentroids

func (km *ElkanClusterer) InitCentroids() error

InitCentroids initializes the centroids using initialization algorithms like random or kmeans++.

func (*ElkanClusterer) SSE

func (km *ElkanClusterer) SSE() float64

SSE returns the sum of squared errors.

type Initializer

type Initializer interface {
	InitCentroids(vectors []*mat.VecDense, k int) (centroids []*mat.VecDense)
}

func NewKMeansPlusPlusInitializer

func NewKMeansPlusPlusInitializer(distFn kmeans.DistanceFunction) Initializer

func NewRandomInitializer

func NewRandomInitializer() Initializer

type KMeansPlusPlus

type KMeansPlusPlus struct {
	// contains filtered or unexported fields
}

KMeansPlusPlus initializes the centroids using kmeans++ algorithm. Complexity: O(k*n*k); n = number of vectors, k = number of clusters Ref Paper: https://theory.stanford.edu/~sergei/papers/kMeansPP-soda.pdf The reason why we have kmeans++ is that it is more stable than random initialization. For example, we have 3 clusters. Using random, we could get 3 centroids: 1&2 which are close to each other and part of cluster 1. 3 is in the middle of 2&3. Using kmeans++, we are sure that 3 centroids are farther away from each other.

func (*KMeansPlusPlus) InitCentroids

func (kpp *KMeansPlusPlus) InitCentroids(vectors []*mat.VecDense, k int) (centroids []*mat.VecDense)

type Random

type Random struct {
	// contains filtered or unexported fields
}

Random initializes the centroids with random centroids from the vector list.

func (*Random) InitCentroids

func (r *Random) InitCentroids(vectors []*mat.VecDense, k int) (centroids []*mat.VecDense)