Mean shift clustering in ML Python - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding the core idea of Mean Shift Clustering

What is the main principle behind the Mean Shift clustering algorithm?

A It partitions data by minimizing the sum of squared distances to cluster centers initialized randomly.

B It builds a hierarchical tree of clusters by merging closest pairs until one cluster remains.

C It assigns points to clusters based on a fixed radius without updating cluster centers.

D It iteratively shifts each data point towards the average of points in its neighborhood to find dense regions.

💡 Hint

Think about how the algorithm finds clusters by moving points towards areas with many neighbors.

Mean Shift works by moving points towards the mean of points in their local neighborhood, effectively climbing the density gradient to find cluster centers.

❓ Predict Output

intermediate

Output of Mean Shift cluster centers

Given the following 2D points and bandwidth, what will be the approximate cluster centers found by Mean Shift?

ML Python

import numpy as np
from sklearn.cluster import MeanShift

points = np.array([[1, 2], [1, 3], [2, 2], [8, 8], [8, 9], [9, 8]])
ms = MeanShift(bandwidth=2)
ms.fit(points)
centers = ms.cluster_centers_
print(centers)

A

[[2.0 3.0]
 [9.0 9.0]]

B

[[1.0 2.0]
 [8.0 8.0]
 [9.0 9.0]]

C

[[1.33333333 2.33333333]
 [8.33333333 8.33333333]]

D

[[1.5 2.5]
 [8.5 8.5]
 [9.5 9.5]]

💡 Hint

Mean Shift averages points close to each other within the bandwidth to find centers.

The first cluster center is the average of points near (1,2) and the second is the average of points near (8,8). The output shows these two centers.

❓ Hyperparameter

advanced

Effect of bandwidth on Mean Shift clustering

What happens if you increase the bandwidth parameter in Mean Shift clustering?

A Clusters become fewer and larger because the neighborhood radius grows, merging more points.

B Clusters become more numerous and smaller because points are grouped more tightly.

C The algorithm runs faster but produces the same number of clusters.

D The algorithm ignores points outside the bandwidth and creates empty clusters.

💡 Hint

Think about how a bigger neighborhood affects grouping points.

Increasing bandwidth means each point considers a larger neighborhood, so clusters merge into bigger groups, reducing the total number.

❓ Metrics

advanced

Choosing a metric to evaluate Mean Shift clustering

Which metric is most appropriate to evaluate the quality of clusters produced by Mean Shift when true labels are unknown?

A Silhouette Score, because it measures how similar points are within clusters compared to other clusters.

B Accuracy, because it compares predicted labels to true labels directly.

C Mean Squared Error, because it measures distance from points to cluster centers.

D Cross-Entropy Loss, because it measures classification error probabilities.

💡 Hint

Think about metrics that work without knowing the true groupings.

Silhouette Score evaluates clustering quality based on cohesion and separation without needing true labels, making it ideal for unsupervised methods like Mean Shift.

🔧 Debug

expert

Debugging Mean Shift convergence issue

Consider this code snippet using Mean Shift. It runs but the cluster centers do not change after the first iteration, causing poor clustering. What is the likely cause?

ML Python

import numpy as np
from sklearn.cluster import MeanShift

points = np.array([[0, 0], [0, 1], [1, 0], [10, 10], [10, 11], [11, 10]])
ms = MeanShift(bandwidth=0.5)
ms.fit(points)
print(ms.cluster_centers_)

A The points array is not normalized, causing the algorithm to fail silently.

B The bandwidth is too small, so points do not have neighbors to shift towards, causing no movement.

C Mean Shift requires labels to be provided, which are missing here.

D The fit method was called on the wrong variable 'ms' instead of 'ms'.

💡 Hint

Check if the bandwidth covers enough points to allow shifting.

A very small bandwidth means each point sees almost no neighbors, so the mean shift step does not move points, resulting in poor clustering.