pre-processing

Question 1

general steps

Answer 1

FeFeSPoT
* transformations (center/scale, skewness, Box-Cox)
* feature extraction
* feature engineering
* predictor selection
* supervised vs unsupervised–supervised considers outcome variable (like PLS)

Question 2

feature extraction

Answer 2

• aka signal extraction
• identifying and extracting features relevant for a particular problem
• eg, PCA (amounting to dimensionality reduction)

Question 3

one-hot encoding / indicator variables

Answer 3

one-hot may refer to encoding every factor level with 0/1, while indicator or dummy variables typically leave one level out (to avoid collinearities)

Question 4

maximum dissimilarity sampling

Answer 4

• a way to stratify test/train sets, by ensuring maximal separation between instances in the predictor phase space
• can also be conditioned on a per-class basis for classification problems

Question 5

resampling types

Answer 5

• LOOCV–for n samples, one is held out for testing, and the other n-1 are used for training
• LGOCV–aka Monte Carlo cross-validation, this just sets a train/test ratio, and then randomly resamples the dataset to create splits on the fly, over some number of repititions
• bootstrap
  
  • random test/train split produced by pulling k samples out of n total instances, with replacement
  • likely there will be samples not picked (due to repeats)–these are the “out of bag” samples, and will be used for testing

Question 6

principal component analysis (PCA)

Answer 6

• works in the per-sample space
• mutually orthogonal linear combinations of predictors that account for the most possible variance
• finds eigenvectors of the predictors’ covariance matrix (which is inherently symmetric); the covariances are over all samples