Hierarchical clustering organizes observations into nested groups, creating a hierarchy rather than requiring a single fixed number of clusters at the outset. This lesson introduces the two main strategies and explains how the hierarchy is represented.
Hierarchical clustering, also called hierarchical cluster analysis (HCA), is a method for building a hierarchy of clusters from data. Nearby or otherwise similar observations are grouped together, and those groups can in turn be combined into larger groups. The result is a nested structure that can be examined at several levels of detail.
All flashcards
What is hierarchical clustering?
A clustering approach that creates nested groups of observations, forming a tree-like hierarchy that can be examined at multiple levels.
What is the starting point of agglomerative clustering?
Every observation starts in its own cluster. The algorithm repeatedly merges clusters until it reaches a stopping condition.
What is the starting point of divisive clustering?
All observations begin in one cluster, which is recursively split into smaller groups.
What does greedy mean in hierarchical clustering?
At each step, the algorithm chooses a locally preferred merge or split according to its criterion, rather than exhaustively comparing every possible hierarchy.
What information does a dendrogram provide?
It displays the nested relationships and sequence of merges or splits. In an agglomerative dendrogram, branch heights usually indicate the dissimilarity at which clusters merge.
Quiz questions
In hierarchical clustering, how do a point metric and a linkage criterion differ?Answer: A point metric compares observations; linkage defines dissimilarity between clusters.
For non-zero vectors, what does cosine distance primarily measure?Answer: Angular dissimilarity, emphasizing vector direction rather than magnitude.
In the DIANA algorithm, what initiates the splinter group?Answer: An object with high average dissimilarity to the rest of its cluster.
How is the symmetric Hausdorff distance between two sets defined?Answer: As the maximum of the two directed set-to-set distances.
How does HAL-x use classifier information when forming its hierarchy?Answer: It merges clusters using classifier-derived estimates of their separability.