Mahout: An overview of clustering techniques

Different kinds of clustering problems

• EXCLUSIVE CLUSTERING In exclusive clustering, an item belongs exclusively to one cluster, not several.
• OVERLAPPING CLUSTERING What if we wanted to do non-exclusive clustering; that is, put Harry Potter not only in fiction but also in a young adult cluster as well as under fantasy. An overlapping clustering algorithm like fuzzy k-means achieves this easily. Moreover, fuzzy k-means also indicates the degree with which an object is associated with a cluster.

 

• HIERARCHICAL CLUSTERING  Now, assume a situation where we have two clusters of books, one for fantasy and the other for space travel. Harry Potter is in the cluster of fantasy books, but these two clusters, space travel and fantasy, could be visualized as subclusters of fiction. Hence, we can construct a fiction cluster by merging these and other similar clusters.

 

• PROBABILISTIC CLUSTERING A probabilistic model is usually a characteristic shape or a type of distribution of a set of points in an n-dimensional plane.

 

Different clustering approaches

• FIXED NUMBER OF CENTERS These clustering methods fix the number of clusters ahead of time.
• BOTTOM-UP APPROACH: FROM POINTS TO CLUSTERS VIA GROUPING

 

• TOP-DOWN APPROACH: SPLITTING THE GIANT CLUSTER