Clustering Algorithm Summary

Recently, we want to build a clustering project on spark, with a large amount of data and classes. The kmeans effect is acceptable, but it is a little slow, so I re-read the common algorithms.

Kmeans
Attention: init Centers (randomize vs kmeans ++)

Mini-Batch kmeans
Loops: random samples; find closest for all; update centers for each

Mean Shift
Init: Get centers by bandwidth
Loops: Find neighbors of centers; update centers; dedup

AP Cluster
Init: Get S; Rik = 0, AIK = 0
Loops: Rik = Sik-max_k '! = K (AIK '+ Sik'); AIK = min (0, rkk + sum_ I '! = I, K max (0, RI 'K); Akk = sum_ I '! = K max (0, RI 'K)
End: For any I, max_k rik + AIK as it's exemplar

Spectral clustering
Steps: similarity Matrix s; S = uv; kmeans of u

Ward hierarchical
Init: each sample as center
Loops: merge to minimize RMSE within clusters

DBSCAN
Init: Get densest core samples
Loops: get more core samples nearby old samples

In fact, scikit-learn implements many algorithms, and there are also ready-made datasets that can be used for experiments. For example: there are some http://scikit-learn.org/stable/modules/clustering.html, and algorithm expansion description.

A comparison of the clustering algorithms in scikit-learn

Method Name	Parameters	Scalability	Usecase	Geometry (metric used)
K-means	Number of clusters	Very largeN_samples, MediumN_clustersWithMinibatch code	General-purpose, even cluster size, flat geometry, not too many ters	Distances between points
Affinity propagation	Damping, sample preference	Not scalable with n_samples	Many clusters, uneven cluster size, non-flat Geometry	Graph distance (e.g. nearest-neighbor graph)
Mean-shift	Bandwidth	Not scalableN_samples	Many clusters, uneven cluster size, non-flat Geometry	Distances between points
Spectral clustering	Number of clusters	MediumN_samples, SmallN_clusters	Few clusters, even cluster size, non-flat Geometry	Graph distance (e.g. nearest-neighbor graph)
Ward hierarchical clustering	Number of clusters	LargeN_samplesAndN_clusters	Many clusters, possibly connectivity constraints	Distances between points
Agglomerative Clustering	Number of clusters, linkage type, distance	LargeN_samplesAndN_clusters	Many clusters, possibly connectivity constraints, Non Euclidean distances	Any pairwise distance
DBSCAN	Neighborhood Size	Very largeN_samples, MediumN_clusters	Non-flat geometry, uneven cluster sizes	Distances between nearest points
Gaussian mixtures	Bytes	Not scalable	Flat geometry, good for Density Estimation	Mahalanobis distances to centers

Clustering Algorithm Summary