P4-17.3 Supplementary Learning: Hierarchical and Spectral Clustering¶

Section ID: P4-17.3 Version: v2026.07.23

After looking at k-means and DBSCAN in P4-17.1, this question naturally remains.

Why are there so many clustering algorithm names, and how are they different from one another?

Two names that come up repeatedly when organizing that question for the first time are hierarchical clustering and spectral clustering.

Rather than going long on implementation details, this supplementary Section distinguishes the two methods from the perspective of what criterion they use to form groups.

Scope Of This Section¶

This Section answers the following questions.

What does hierarchical clustering show step by step?
Why is spectral clustering said to look at connectivity before distance?
Compared with k-means and DBSCAN, in what scenes do these two methods come to mind?
When distinguishing the four methods for the first time, what is the first thing worth recording?

This Section focuses on distinguishing the perspective of merging like a tree from the perspective of cutting and grouping a graph.

Goals Of This Section¶

You can explain hierarchical clustering as clustering that shows the order in which things merge.
You can explain spectral clustering as clustering that regroupes data by rereading connectivity structure.
You can describe that k-means, DBSCAN, hierarchical clustering, and spectral clustering start from different questions.

Why This Section Is Needed¶

When first learning clustering, the growing list of algorithm names can make things more confusing rather than less.

k-means looks at centers
DBSCAN looks at density
But then some methods are described as tree-like, while others are described as graph-like

The important thing here is not memorizing more names. It is to first grasp what each algorithm is trying to use to read clusters.

So P4-17.3 reorganizes the clustering family through four intuitions: center, density, merge order, and connectivity.

What Does Hierarchical Clustering Show?¶

Hierarchical clustering does not begin by cutting points into a fixed number of clusters from the start. Instead, it tries to show step by step what becomes close first and what merges later.

At an introductory level, one sentence is enough.

Hierarchical clustering shows not only one clustering result, but also the order in which groups grow.

Suppose there are four customers.

A and B are very similar
C and D are also fairly similar
The A/B group and the C/D group are somewhat farther from each other

In that case, hierarchical clustering can show the order in which A and B merge first, C and D merge next, and then the two groups merge into a larger group at the end.

flowchart TB
  A["A"]
  B["B"]
  C["C"]
  D["D"]
  AB["A+B"]
  CD["C+D"]
  ALL["(A+B)+(C+D)"]

  A --> AB
  B --> AB
  C --> CD
  D --> CD
  AB --> ALL
  CD --> ALL

The key point in this diagram is the merge order, not the final number of clusters. So hierarchical clustering naturally comes to mind when you have not decided yet how many groups to cut into, but you want to see the structure of what becomes close first.

This difference can be stated more briefly like this.

What you want to know first	Why hierarchical clustering fits better
It is still unclear whether to read the data as 2 groups or 4 groups	Because you can look at the merge order and decide the cut point later
You want to see which samples become similar to each other first	Because it shows the stages of becoming close, not only the final number of clusters
You want to adjust the grouping level together with domain review	Because it lets you inspect multiple grouping levels instead of only one fixed answer

So hierarchical clustering is closer to first looking at how groups grow, then deciding later at what height to cut and read them than to finding the correct number of clusters in one shot. That is why it is especially useful in early exploration, when domain interpretation is not settled yet.

What Is Spectral Clustering Trying To See?¶

Spectral clustering does not try to read points only through coordinate distance. It tries to reread them through a graph structure that asks who is connected to whom.

For an introduction, the following is enough.

Spectral clustering tries to read breaks in connectivity better than the shape of nearby points alone.

Suppose there are two curved groups shaped like crescents. In a scene like this, center-based k-means may cut the boundary too straight, but if you look at connectivity structure, it can be more natural to read points on the same curve as one group.

flowchart TD
  A["points"]
  B["build similarity graph"]
  C["find connected structure"]
  D["split graph into groups"]

  A --> B --> C --> D

The key here is not do we cut the original coordinates directly? but do we build a new expression of the connectivity relation first, then split that structure? So spectral clustering comes to mind in scenes where the shape is complex, but the connectivity is still clear.

This can be compared more directly as follows.

The first problem you hit	Why spectral clustering comes to mind
A center-based cut breaks a curved structure in the middle	Because it is more natural to read connectivity before geometric centers
Some points look far apart, but still continue along the same curve	Because a graph of neighborhood relations can reveal connectivity inside the same structure more clearly
The cluster shape does not look like round blobs	Because rebuilding the graph structure may fit better than relying on one distance alone

Still, spectral clustering is not always the right answer just because the shape is complex. The core question is whether building a graph from neighborhood relations reveals the connectivity inside the same structure more clearly. In other words, it is more accurate to understand spectral clustering as a way of rerepresenting connectivity and then reading that structure, rather than directly cutting coordinates.

Putting The Four Intuitions Side By Side¶

If we place the representative intuitions from Chapter 17 into one table, they organize like this.

Method	What it looks at first	Scene where it naturally comes to mind	What to be careful about first
k-means	center	When you want to quickly see round and relatively even groups	It is sensitive to the assumed number and shape of clusters
DBSCAN	density	When you want to separate noise and inspect irregular groups	It is sensitive to `eps` and `min_samples`
Hierarchical clustering	merge order	When you want to see how groups grow step by step	You still need to interpret where to cut
Spectral clustering	connectivity	When the shape is complex but the connectivity structure is clear	It is sensitive to how the similarity graph is built

The purpose of this table is not ranking the methods. The first thing is to hold on to the fact that they read clusters through different handles.

If this comparison is turned into concrete selection questions, it looks like this.

The question you ask first right now	The method that comes to mind first
Do you want to see the merge order more than fix the number of clusters immediately?	Hierarchical clustering
Do you want to keep noise separate and inspect irregular groups?	DBSCAN
Do you want a quick baseline for round and relatively even groups?	k-means
Do you want to cut a curved or complex connectivity structure?	Spectral clustering

If we write the difference between hierarchical clustering and spectral clustering more directly, it looks like this.

The question you hit first	The more natural starting point
Are you more interested in `what merges first` than in `how many groups to cut into`	Hierarchical clustering
Is the order of getting closer itself a clue for interpretation	Hierarchical clustering
Is there a curved structure that center points fail to capture well	Spectral clustering
Does `who is connected to whom` seem more important than coordinate distance	Spectral clustering

Cases And Examples¶

Case 1. When You Have Not Yet Decided At What Level To Group Customer Segments¶

Suppose a subscription service team wants to group customers, but it is difficult to fix the analysis from the start as 3 clusters or 4 clusters. In that case, hierarchical clustering can be a useful starting point. By seeing which customers become close first and at what stage they merge into larger groups, the team can read is it more natural to see this as 2 groups right now, or as 4 groups together with domain review.

So in this scene, the important thing is not automatically finding the correct number of clusters. It is to look at the order in which groups grow and build interpretation candidates more carefully.

flowchart TD
  A["customer behavior table"]
  B["fixed k is still unclear"]
  C["inspect which groups merge first"]
  D["compare 2-group and 4-group cuts"]
  E["review segment meaning with domain context"]

  A --> B --> C --> D --> E

Case 2. When Curved Movement Paths Are Hard To Cut Well With Centers¶

When data take a long curved shape, such as location logs or movement trajectories, k-means may cut them too straight based on centers. In such a scene, there are cases where you want to care more about points connected along the same curve, and that is when spectral clustering can come to mind. In other words, it is a problem scene where does this stay connected inside the same structure matters more than what is the nearest center.

flowchart TD
  A["curved trajectory points"]
  B["center-based cut looks awkward"]
  C["inspect neighborhood connections"]
  D["split by connected structure"]
  E["review whether the curve stays intact"]

  A --> B --> C --> D --> E

Looking at the two cases together makes the difference clearer.

The question read first in the case	The method that fits better	Why
Is it still unclear at what grouping level customer segments should be read	Hierarchical clustering	Because you can compare multiple cut levels while looking at the merge order
Do you want to see whether curved movement paths stay inside the same structure	Spectral clustering	Because it is more natural to read connectivity before centers

In short, hierarchical clustering is closer to interpreting the order in which groups grow, while spectral clustering is closer to inspecting whether connectivity structure is preserved.

Practice And Example¶

This practice block is a verification exercise: choose first what clustering intuition should come to mind and then compare your choice with the explanation right below.

Problem situation: distinguish, scene by scene, whether round groups, irregular groups, merge order, or connectivity structure matters first
Input: four short problem scenes
Expected output: for each scene, the method that should come to mind first and the reason
Concepts to check:
There are questions that must be read before algorithm names
Even with the same data, the starting point changes depending on what structure you want to inspect

First cover the Method to think of first column, write down your own choice for each scene, and then compare it with the table.

Scene	Method to think of first	Why this method fits first
You want a quick baseline for round and fairly even groups	k-means	Because the question starts with putting centers down and grouping quickly
You want to leave noise aside and inspect irregular groups	DBSCAN	Because density and noise separation matter first in this scene
You want to see what merges first before fixing the number of clusters	Hierarchical clustering	Because merge order itself is the interpretation clue here
You want to preserve a curved structure, and a center-based cut looks awkward	Spectral clustering	Because reading connectivity first is more natural than reading centers first

The point of this practice is not memorizing algorithm names. It is to ask first what structure do I want to inspect? More important than getting the label right is being able to explain why that question comes first in the scene.

What Is Worth Recording In A First Comparison?¶

At the beginning, it is more practical to record the following questions than to focus on implementation details.

Question to record	Why it is needed
Do we need to fix the number of clusters from the start	Because this becomes the first criterion for separating k-means from other methods
Do we need to keep noise separate	Because this becomes the criterion for whether DBSCAN-like intuition is needed
Does the merge order itself matter	Because this is a reason to think of hierarchical clustering
Is the shape complex, but does connectivity seem more important	Because this is a reason to think of spectral clustering

If you record these questions first, you will read more easily what clustering intuition the current data scene is asking for, instead of trying to memorize algorithm names.

Checklist¶

Can you explain that hierarchical clustering shows what merges first more than how many groups there are?
Do you understand that spectral clustering is closer to rereading connectivity structure than to rereading distance itself?
Is this a scene where you want to see what becomes close first, rather than fixing the number of clusters immediately?
Is this a scene where connectivity structure must be read first, not merely one with an irregular shape?
Instead of reading the clustering result as one final answer, can you explain from what question this method came to mind?
Can you distinguish k-means, DBSCAN, hierarchical clustering, and spectral clustering again through the four axes of center, density, order, and connectivity?

Sources And References¶

scikit-learn developers, 2.3. Clustering, scikit-learn User Guide. Consulted for the basic descriptions and comparison axes of hierarchical clustering, spectral clustering, k-means, and DBSCAN. Accessed 2026-07-19. https://scikit-learn.org/stable/modules/clustering.html
Ulrike von Luxburg, A Tutorial on Spectral Clustering, Statistics and Computing, 2007. Consulted for the background connecting spectral clustering to graph Laplacians and connectivity structure. Accessed 2026-07-19. https://doi.org/10.1007/s11222-007-9033-z