Hierarchical clustering in ML Python - Model Metrics & Evaluation

Metrics & Evaluation - Hierarchical clustering

Which metric matters for Hierarchical clustering and WHY

Hierarchical clustering groups data points based on similarity without predefined labels. Because it is unsupervised, common metrics like accuracy do not apply. Instead, we use cophenetic correlation coefficient to measure how well the clustering preserves the original distances between points. Another key metric is the silhouette score, which shows how well each point fits within its cluster compared to other clusters. These metrics help us understand if the clusters are meaningful and well-separated.

Confusion matrix or equivalent visualization

Hierarchical clustering does not produce a confusion matrix because it is unsupervised and has no true labels. Instead, we visualize results using dendrograms and silhouette plots.

Dendrogram example:

  |          _______
  |         |       |
  |    ____ |       |____
  |   |    ||            |
  |___|____||____________|____

This tree-like diagram shows how clusters merge step-by-step.

Silhouette plot example:

  Silhouette score ranges from -1 to 1.
  Higher values mean better clustering.

  |  *  *  *  *  *  *  *  *  *  *
  |  *  *  *  *  *  *  *  *  *  *
  |  *  *  *  *  *  *  *  *  *  *
  |______________________________

Precision vs Recall tradeoff with concrete examples

In hierarchical clustering, precision and recall do not directly apply because there are no true labels. Instead, the tradeoff is between cluster tightness and cluster separation.

• Tight clusters mean points inside a cluster are very similar (high cohesion), but this can cause many small clusters (over-segmentation).
• Well-separated clusters mean clusters are distinct from each other, but forcing separation can group dissimilar points together (under-segmentation).

Choosing the right level to cut the dendrogram balances these. For example, in customer segmentation, too many small groups may be hard to act on, while too few groups may hide important differences.

What "good" vs "bad" metric values look like for Hierarchical clustering

Good values:

• Cophenetic correlation coefficient close to 1 (e.g., > 0.8) means the clustering preserves original distances well.
• Silhouette score near 0.5 or higher indicates clear, well-separated clusters.

Bad values:

• Cophenetic correlation below 0.5 means poor representation of data structure.
• Silhouette score near 0 or negative means clusters overlap or points are assigned poorly.

Metrics pitfalls in Hierarchical clustering

• Ignoring scale: Features with different scales can distort distance calculations, hurting cluster quality.
• Overfitting dendrogram cuts: Cutting too deep creates many tiny clusters that may not be useful.
• Using silhouette score blindly: It can be misleading if clusters have very different sizes or shapes.
• Data leakage: Using labels or future information to evaluate clusters breaks unsupervised assumptions.

Self-check question

Your hierarchical clustering model has a cophenetic correlation coefficient of 0.9 but a silhouette score of 0.2. Is this good?

Answer: The high cophenetic correlation means the clustering preserves the original distances well, which is good. However, the low silhouette score indicates clusters are not well separated or points may be assigned poorly. This suggests the clusters may not be meaningful or distinct enough. You might need to adjust the clustering method or how you cut the dendrogram to improve cluster quality.

Key Result

For hierarchical clustering, cophenetic correlation and silhouette score are key to judge cluster quality and separation.

Practice

1. What is the main idea behind hierarchical clustering in machine learning?

easy

A. It uses a neural network to classify data points into clusters.

B. It assigns data points to clusters based on a fixed number of clusters.

C. It builds a tree of clusters by joining or splitting data points step by step.

D. It randomly assigns data points to clusters without any order.

Solution

1. Step 1: Understand hierarchical clustering concept

   Hierarchical clustering creates a tree-like structure by either merging or splitting data points stepwise.

2. Step 2: Compare with other clustering methods

   Unlike fixed cluster methods, it does not require a preset number of clusters and builds a hierarchy instead.

3. Final Answer:

   It builds a tree of clusters by joining or splitting data points step by step. -> Option C

4. Quick Check:

   Hierarchical clustering = stepwise tree building [OK]

Hint: Think of building a family tree of data points step by step [OK]

Common Mistakes:

• Confusing hierarchical clustering with k-means clustering
• Thinking it requires a fixed number of clusters upfront
• Believing it assigns clusters randomly

2. Which of the following Python code snippets correctly imports the function for hierarchical clustering from the SciPy library?

easy

A. from scipy.cluster.hierarchy import linkage

B. import sklearn.cluster.hierarchical as hc

C. from scipy.cluster import kmeans

D. import hierarchical from scipy.cluster

Solution

1. Step 1: Identify correct import for hierarchical clustering

   The function linkage from scipy.cluster.hierarchy is used for hierarchical clustering.

2. Step 2: Check other options for errors

   A has invalid syntax, B uses wrong module name (sklearn.cluster.hierarchical does not exist), C imports kmeans (not hierarchical).

3. Final Answer:

   from scipy.cluster.hierarchy import linkage -> Option A

4. Quick Check:

   Correct import = from scipy.cluster.hierarchy import linkage [OK]

Hint: Remember SciPy's hierarchy module has linkage function [OK]

Common Mistakes:

• Using sklearn.cluster.hierarchical which does not exist
• Confusing kmeans import with hierarchical clustering
• Incorrect import syntax

3. Given the following Python code using SciPy for hierarchical clustering, what will be the output of print(Z)?

from scipy.cluster.hierarchy import linkage
import numpy as np

X = np.array([[1, 2], [1, 4], [5, 2], [5, 4]])
Z = linkage(X, method='single')
print(Z)

medium

A. A 3x4 array showing the order and distance of cluster merges

B. A 4x4 distance matrix of all points

C. A list of cluster labels for each point

D. An error because method 'single' is invalid

Solution

1. Step 1: Understand linkage output format

   The linkage function returns a matrix with rows showing which clusters merged, their distance, and cluster size.

2. Step 2: Check method and output shape

   Method 'single' is valid. For 4 points, output is (4-1)=3 rows and 4 columns.

3. Final Answer:

   A 3x4 array showing the order and distance of cluster merges -> Option A

4. Quick Check:

   linkage output = cluster merges matrix [OK]

Hint: linkage returns merge info, not labels or distance matrix [OK]

Common Mistakes:

• Expecting cluster labels instead of merge matrix
• Confusing linkage output with distance matrix
• Thinking 'single' method is invalid

4. You run this code but get an error:

from scipy.cluster.hierarchy import dendrogram
import numpy as np

X = np.array([[1, 2], [3, 4], [5, 6]])
dendrogram(X)

What is the problem and how to fix it?

medium

A. The array X has wrong shape; reshape it to 1D before dendrogram().

B. You must pass a linkage matrix, not raw data, to dendrogram(). Use linkage() first.

C. dendrogram() requires labels argument; add labels=['a','b','c'].

D. You need to import dendrogram from sklearn.cluster, not scipy.

Solution

1. Step 1: Identify dendrogram input requirements

   dendrogram() expects a linkage matrix, not raw data points.

2. Step 2: Fix by computing linkage first

   Use linkage() on X to get the linkage matrix, then pass it to dendrogram().

3. Final Answer:

   You must pass a linkage matrix, not raw data, to dendrogram(). Use linkage() first. -> Option B

4. Quick Check:

   dendrogram input = linkage matrix [OK]

Hint: Always run linkage() before dendrogram() [OK]

Common Mistakes:

• Passing raw data directly to dendrogram()
• Thinking dendrogram needs labels argument
• Importing dendrogram from wrong library

5. You want to cluster customer data using hierarchical clustering but do not know how many clusters to choose. Which approach helps you decide the number of clusters without setting it upfront?

hard

A. Use PCA to reduce dimensions before clustering.

B. Use k-means clustering instead with a fixed k value.

C. Randomly assign clusters and check accuracy.

D. Cut the dendrogram at a chosen distance threshold to form clusters.

Solution

1. Step 1: Understand hierarchical clustering flexibility

   Hierarchical clustering builds a tree and allows cutting at any distance to form clusters.

2. Step 2: Use distance threshold to decide clusters

   By choosing a height to cut the dendrogram, you get clusters without presetting their number.

3. Final Answer:

   Cut the dendrogram at a chosen distance threshold to form clusters. -> Option D

4. Quick Check:

   Distance threshold cutting = flexible cluster count [OK]

Hint: Cut dendrogram by distance, not fixed cluster count [OK]

Common Mistakes:

• Thinking hierarchical clustering needs fixed cluster count upfront
• Confusing with k-means clustering method
• Ignoring dendrogram visualization for cluster choice