Module 6

Question 1

K-means

Answer 1

• select K
• generate K random cluster centroids
• assign each training example to the nearest centroid
• update position of each centroid
• stop if position of centroids didn’t change otherwise go to step 2

Question 2

Elbow method

Answer 2

• run K-means multiple times with different K’s
• keep track of cost L(x) for each K value
• select K where the rate of decrease sharply shifts

Question 3

Cross-validation to select K

Answer 3

• split dataset in N folds
• N-1 folds off compute centroids positions with k-means
• compute average score on validation datasets

Question 4

K-means pros

Answer 4

• easy to understand/implement
• used very often for clustering
• efficient

Question 5

K-means cons

Answer 5

• have to define K
• local optimum is sensitive to the initial centroid positions
• not suitable to discover clusters that aren’t hyper-ellipsoids

Question 6

GMM-EM

Answer 6

• select K, initialise all parameters
• compute the responsibilities
• update the mean
• update the covariance
• update the mixing proportions
• stop if converged, otherwise go to step 2