ML-07 - Unsupervised learning

Question 1

ML-07 - Unsupervised learning

What’s the requirement for unsupervised learning?

Answer 1

Your data needs some inherent structure.

Question 2

ML-07 - Unsupervised learning

What are the two major types of unsupervised learning?

Answer 2

• Clustering
• Anomaly detection

Question 3

ML-07 - Unsupervised learning

Name the techniques mentioned in the lecture slides. (5)

Answer 3

• K-Means Clustering
• Mean-Shift Clustering
• DBSCAN
• EM using GMM
• Agglomerative Hierarchical Clustering

Question 4

ML-07 - Unsupervised learning

What is DBSCAN short for?

Answer 4

Density-Based Spatial Clustering Applications with Noise

Question 5

ML-07 - Unsupervised learning

What is EM short for?

Answer 5

Expectation maximixation

Question 6

ML-07 - Unsupervised learning

What is GMM short for?

Answer 6

Gaussian Mixture Models

Question 7

ML-07 - Unsupervised learning

What is “EM using GMM” short for?

Answer 7

xpectation Maximization Clustering using Gaussian Mixture Models

Question 8

ML-07 - Unsupervised learning

How do you perform k-means clustering?

Answer 8

1) Assign N centroids to your data at random.
2) Assign data point to nearest centroid.
3) Move centroid to average of data positions.
4) Repeat 2+3 until convergence.

Question 9

ML-07 - Unsupervised learning

How do you select the number of clusters to use in k-means?

Answer 9

• Eyeball data (if few dimensions)
• Elbow method
• Analyze within-cluster variance

Question 10

ML-07 - Unsupervised learning

Describe the elbow method.

Answer 10

1) Plot cost vs. n_clusters
2) Look for the “elbow”, i.e. where the cost decrease slows down.

(See image)

Question 11

ML-07 - Unsupervised learning

When does the elbow method work the best?

Answer 11

When you have separated clusters.

(See image)

Question 12

ML-07 - Unsupervised learning

When does the elbow method not work that well?

Answer 12

If you have non-separated clusters with uncertain boundaries. There might not be an elbow in the plot.

Question 13

ML-07 - Unsupervised learning

What are the pros of using k-means?

Answer 13

Fast computation

Question 14

ML-07 - Unsupervised learning

What are the drawbacks of k-means?

Answer 14

• Challenging to identify the groups/clusters.
• Random initialization means the results can lack consistency.

Question 15

ML-07 - Unsupervised learning

What two classes of methods are used to evaluate clustering performance?

Answer 15

• Supervised
• Unsupervised

Question 16

ML-07 - Unsupervised learning

What is supervised clustering evaluation?

Answer 16

Compare predicted vs. actual labels.

Question 17

ML-07 - Unsupervised learning

What is the silhouette coefficient?

Answer 17

An unsupervised clustering evaluation technique.

Question 18

ML-07 - Unsupervised learning

Is the silhouette coefficient a supervised or unsupervised method?

Answer 18

Unsupervised.

Question 19

ML-07 - Unsupervised learning

How does the silhouette coefficient work?

Answer 19

It scores the clustering model’s performance by checking that:
- Inter-cluster distances should be maximized.
- Inter-cluster distances should be minimized

Question 20

ML-07 - Unsupervised learning

What is the formula for the silhouette coefficient?

Answer 20

(See image)

The mean value of all samples is then used.

Question 21

ML-07 - Unsupervised learning

How do you evaluate the results of the silhouette coefficient?

Answer 21

• -1 is bad
• 1 is good.
• 0 indicates overlapping clusters.