Normality Tests

• testing whether sample-data has a normal distribution

Transforming Data to Normal Distribution

• transforming a data variable to take on a normal distribution

Feature Engineering

is the process of using domain knowledge to extract features from raw data and creating feature functions

Normalization

given 𝒙 = {𝑥₁, …, 𝑥_𝑛}

• min-max normalization(𝑥ᵢ) = [𝑥ᵢ - 𝑚𝑖𝑛(𝒙)] / [𝑚𝑎𝑥(𝒙) - 𝑚𝑖𝑛(𝒙)]
• mean normalization(𝑥ᵢ) = [𝑥ᵢ - 𝜇] / [𝑚𝑎𝑥(𝒙) - 𝑚𝑖𝑛(𝒙)]
• z-score normalization(𝑥ᵢ) = [𝑥ᵢ - 𝜇] / 𝜎 # transform all variables to have same standard deviation

where:

• 𝜇 - mean
• 𝜎 - standard deviation

Detecting Outliers

• removing and detecting outliers that skew the model

Discretization

• divides the range of a continuous attribute into intervals
• equal-width (distance) binning
  • 𝑏𝑖𝑛-𝑤𝑖𝑑𝑡ℎ = (𝑚𝑎𝑥 - 𝑚𝑖𝑛) / 𝐾
• equal-depth (frequency) binning
  • divides the range intervals each containing the same number of data points
• bottom-up binning
  • criteria such as entropy to characterize the purity of bins
• clustering

Nominal & Ordinal

sometimes we want to convert nominal and ordinal values to “continuous” values

• nominal - values from an unordered set (colors, profession, etc.)
• ordinal - values from an ordered set (rank, etc.)

Missing Data

• Ignoring Instances with Unknown Feature Values
• Most Common Feature Value
• Mean substitution
• Regression or Classification methods
• Nearest-Neighbor Imputation
• Treating Missing Feature Values as Special Values
• Latent Variable Methods