lecture 13 NMF

Question 1

PCA

Answer 1

Orthogonal
Sparse pca: components orthogonal& sparse

Question 2

Nmf

Answer 2

Latent representation and latent features are non- negative

Question 3

Nmf

Answer 3

Easier to interpret
No cancellation like pca
No sign ambiguity
Can learn over complete representation
Soft clustering

Question 4

Downside of Nmf

Answer 4

Only applied to non- negative
Interpretable is hit/ miss
Non- convex optimization
Slow on large dataset
Not orthogonal

Question 5

Manifold learning

Answer 5

Allow more complex transformations
Better visulizatjon -> less than 2
New representation of training not test
Good for eda

Question 6

Tsne

Answer 6

Random 2D
Points - close
Far——-far
Emphasis on points that are close by

Question 7

Outliers— parametric

Answer 7

Elliptic — Gaussian model— fit robust covariants matrix and mean

Only works if Gaussian assumption is reasonable

Question 8

Kernel density

Answer 8

Non- parametric model
Need to adjust kernel bandwidth
Not good in high dimension

Question 9

One class svm

Answer 9

Use Gaussian kernel to cover data
Select support vectors
Select gamma
Specify outlier ratio

Question 10

Isolation tree

Answer 10

Outliers are easier to isolate from the rest