Clustering

Clustering

Clustering of unlabeled data can be performed with the module sklearn.cluster. Each clustering algorithm comes in two variants: a class, that implements the fit method to learn the clusters on train data, and a function, that, given train data, returns an array of integer labels corresponding to the different clusters. For the class, the labels over the training data can be found in the labels_ attribute.

in summary, methods:

• - based on parameters:

  n_clusters as params: K-Means, Spectral, ward hierarchical, agglomerative, OPTICS (minimum)
  neighborhood size: DBSCAN
  many params: Gaussian Mixtures
  distance threshold: ward hierarchical, agglomerative, birch
  bandwidth: Mean-shift
  sample preferences/damping: affinity propagation

• - based on scalability

  large n_samples, medium n_clusters: K-Means, DBSCAN
  medium n_samples, small n_clusters: spectral
  large n_samples, large n_clusters: ward hierarchical, agglomerative, OPTICS, Birch*
  not-scalable with n_samples: affinity propagation, Mean-shift, Guassian Mixture

  * note that Birch is used in Large datasets, for outlier removal and data reduction.

• - based on n_clusters

  many clusters: affinity propagation, Mean-shift, ward hierarchical, agglomerative, medium clusts: K-Means, spectral

• - based on geometry

  flat: K-Means, Gaussian Mixture
  non-flat: affinity propagation, Mean-shift, spectral, DBSCAN, OPTICS

• - based on metric used

  dist between pooint: K-Means, Mean-shift, ward hierarchical, DBSCAN, OPTICS, Birch, Gaussian Mixtures
  graph distance: affinity propagation, Spectral
  any pairwise dist: agglomerative

• - if distance threshold or neighborhood size methods considered:

  ward hierarchical, agglomerative, K-Means, Mean-shift, DBSCAN, OPTICS, Birch, Gaussian Mixtures

• - if graph distance method considered:

  affinity propagation, Spectral

As an another view:

if flat geometry considered: (all not scalable with n_samples)

1- if many n_clusters considered:

• if distance threshold method: use Mean-shift
• if graph distance method: use affinity propagation

2- if no matter how many n_clusters: use Gaussian Mixtures (distance threshold method)

if non-flat geometry considered:

if large n_samples and matter how many n_clusters: (all using distance threshold method)

• if many: use ward hierarchical, agglomerative, OPTICS, Birch
• if medium/few: use K-Means, DBSCAN

if medium n_samples: use Spectral (using graph distance method)

Check this page for detailed information and comparison between methods.

>

Have Fun!