26_hierarchical_clustering
26. Hierarchical Clustering
1. Overview
Hierarchical clustering is an agglomerative (bottom-up) approach that builds a nested tree of clusters, known as a dendrogram. Unlike K-Means, it does not require the user to pre-specify the number of clusters K prior to execution.
2. The Agglomerative Algorithm
- Initialization: Begin by treating every single observation as its own independent cluster (if there are n observations, start with n clusters).
- Iterative Merging:
- Identify the two clusters that are most similar (i.e., have the shortest distance between them).
- Merge these two clusters into a single new cluster.
- Termination: Repeat the merging process until all observations are encapsulated within a single, massive root cluster.
3. Linkage Criteria
To determine the distance between two clusters (which may contain multiple points), a linkage criterion is required. Let A and B be two clusters:
- Single Linkage: The distance between the closest pair of points (one in A, one in B). Tends to produce long, trailing clusters (chaining) where observations are fused one at a time.
\min_{i \in A, j \in B} d(x_i, x_j)
- Complete Linkage: The distance between the farthest pair of points. Tends to produce compact, balanced, spherical clusters.
\max_{i \in A, j \in B} d(x_i, x_j)
- Average Linkage: The arithmetic mean of distances between all pairs of points. Also tends to yield more balanced dendrograms.
- Centroid Linkage: The distance between the calculated centroids of clusters A and B. Often used in genomics, but suffers from a major drawback called inversion, whereby two clusters can fuse at a height below either of the individual clusters in the dendrogram.
4. The Dendrogram
The algorithm outputs a dendrogram, a tree diagram where the vertical axis represents the distance/dissimilarity at which clusters were merged.
- Choosing K: The user can obtain a specific number of clusters by drawing a horizontal line across the dendrogram at a chosen height (cutting the tree).
- Interpretation (Critical Rule): The height of the merge indicates how different the two groups are; higher merges indicate joining highly dissimilar groups. It is incorrect to draw conclusions about the similarity of two observations based on their proximity along the horizontal axis. There are 2^{n-1} possible reorderings of the dendrogram branches. Similarity is exclusively determined by the vertical location where branches first fuse.
- Nesting Limitation: Hierarchical clustering forces clusters to be nested. If the true subgroups are not nested (e.g., the best 2-group split is by gender, but the best 3-group split is by nationality), hierarchical clustering will yield poor results compared to K-Means.
5. Dendrogram Structure
Dissimilarity
^
High | [---- Cluster A&B&C ----]
| | |
| [A&B] |
Low | | | |
| (A) (B) (C)
+-------------------------------------> Data Points
Use Cases & Limitations
5 Use Cases (When to use):
- Unknown Cluster Count: Ideal for exploratory data analysis when you have absolutely no idea how many clusters exist (K is not required upfront).
- Taxonomy & Evolutionary Trees: Naturally suited for bioinformatics (genetics/phylogenetics) to map out evolutionary descent in a strict hierarchy.
- Visual Data Structure Discovery: The resulting dendrogram provides a highly interpretable visual map of exactly how all subgroups relate to one another at varying levels of granularity.
- Small Datasets: Highly effective on smaller datasets where the rich visual output can be comprehensively analyzed by domain experts.
- Document Organization: Useful for categorizing documents where natural hierarchies exist (e.g., Science -> Physics -> Quantum Mechanics).
5 Limitations (When NOT to use):
- Massive Computational Expense: The standard algorithm has a time complexity of O(n^3) and a space complexity of O(n^2), making it entirely unusable for large datasets.
- Irreversible Merges: Once two clusters are merged at a lower level in the tree, they can never be separated later, even if subsequent steps reveal it was a suboptimal merge.
- Non-Nested True Structures: If the true underlying data segments are not naturally nested (e.g., splitting by gender vs. splitting by country), forcing a hierarchical structure yields misleading results.
- Outlier Sensitivity: Highly sensitive to noise and outliers, particularly when using "Single Linkage," which can result in severe "chaining" where outliers are appended one by one.
- Subjective Tree Cutting: While you don't need to pick K initially, determining where to draw the horizontal line across the dendrogram to establish the final clusters remains highly subjective.