SHORT QUESTIONS FOR HIERARCHICAL CLUSTERING
General Introduction.
Hierarchical clustering is a type of unsupervised machine learning algorithm used to group
similar objects into clusters based on their distance or similarity. It creates a hierarchy of clusters
that can be visualized using a dendrogram (a tree-like diagram). The process can be:
Agglomerative (Bottom-Up): Starts with each data point as its own cluster and merges
the closest clusters iteratively.
Divisive (Top-Down): Starts with all data points in one cluster and splits them
iteratively.
Q1. When should you use hierarchical clustering, and when not?
Use hierarchical clustering when:
You want to understand nested groupings or structure in data.
You're working with small to medium-sized datasets.
You need to visualize clustering with a dendrogram.
Avoid it when:
You're dealing with very large datasets (due to performance issues).
You need real-time or very fast clustering.
You have high noise or irrelevant features (it can skew the results).
Use hierarchical clustering when interpretability is more important than speed.
Q2. What are the advantages and disadvantages of hierarchical clustering compared to
other methods like K-means?
Advantages:
No need to pre-specify the number of clusters.
Produces a hierarchy, giving deeper insight into data structure.
Dendrogram helps visualize the process.
Disadvantages:
Computationally expensive: Not suitable for very large datasets (O(n²) complexity).
Sensitive to noise and outliers.
Once a merge or split is made, it cannot be undone (greedy algorithm).
Comparison with K-means:
K-means needs the number of clusters in advance; hierarchical doesn’t.
K-means is faster and scalable, but can miss complex relationships.
Hierarchical is better when data relationships are nested or not well-separated.
Q3. What are the two main types of hierarchical clustering, and how do they compare?
There are two primary types:
A. Agglomerative (Bottom-Up) Clustering:
Each data point starts as its own cluster.
At every step, the closest two clusters are merged.
This continues until all points are merged into a single large cluster (the root of the
hierarchy).
Most commonly used.
B. Divisive (Top-Down) Clustering:
Starts with one large cluster containing all data points.
At each step, the least similar data points are split off to form new clusters.
This continues until each point is in its own cluster.
Less common due to higher complexity.
Comparison:
Agglomerative is like building a puzzle piece-by-piece.
Divisive is like breaking a picture apart into its components.
Q4. What kinds of data or problems are best suited for hierarchical clustering?
Hierarchical clustering shines when the data has structure that is nested, gradual, or tree-like.
If your data has layers of meaning or organization — like taxonomies, evolutionary
trees, or document topics — hierarchical clustering can reveal relationships you didn’t
even know to look for.
It’s ideal for exploratory analysis, where the goal is insight, not just quick
segmentation.
Examples:
Gene expression data: Understand how genes group together based on similar patterns.
Document clustering: Reveal topic hierarchies in texts (e.g., politics → elections →
candidates).
Customer behavior: Discover if groups of customers form subgroups with shared habits.
Q5. How does the choice of distance metric and linkage method influence hierarchical
clustering results, and which linkage method is best?
Ans. The distance metric and linkage method significantly impact the results of hierarchical
clustering. Different combinations, like Euclidean distance with complete linkage versus cosine
distance with average linkage, can yield completely different groupings. This is crucial because
the choice influences both the stability and interpretability of the clusters. For example, single
linkage might produce elongated, chain-like clusters, while complete linkage tends to create
more compact, spherical clusters. There’s no “one-size-fits-all” best linkage method — it
depends on the data and the intended interpretation. Ward’s method is often preferred when
aiming for balanced clusters with minimal variance, making it a solid choice for many practical
applications, especially when dealing with continuous data. However, choosing the right method
requires balancing computational efficiency with the specific nature of the data being analyzed.
