Regression and Clustering

Question 1

Ridge

Answer 1

Shrinks coefficients towards 0 but never exactly 0.

Question 1

Lasso (L1)

Answer 1

Some are exactly 0

• Like embedded feature selection.

Question 2

Regularisation

Answer 2

if $w$ is not REGULARISED. THEN they can explode. And you get overfitting. Regularisation is a penalty to keep $w$ under control.

Question 3

K-Means steps

Answer 3

1. Start with k random cluster centres/centroids
2. Assign each object to the nearest centroids
3. Compute the new centroid for each cluster as the mean of the objects assigned to the cluster
4. Repeat step 2 until no change to the centroids

Question 4

K-Means pros

Answer 4

Advantages:

• Simple
• Flexible
• Scales well to a large dataset(features and samples)

Question 5

K-Means cons

Answer 5

Disadvantages:

• Have to specify K
• Categorical data
• Need to re-run to obtain clustering for different k
• Stochastic (non-deterministic). Having different starting centroids will produce different results.
• Usually convert to local optima