Preprocessing Numerical Datasets

1. Review data quality: Identify missing values, eliminate columns or rows with missing values. Simple imputation: use mean, mode, or median to fill in missing values. Advanced imputation: use KNN or regression to fill in missing values.
2. Check for duplicate rows.
3. Identify outliers: Values that fall outside the typical data range.
   • Standard deviation
   • Interquartile range
4. Data scaling (Normalization): Theory suggests standardizing data between 0 and 1.
5. Logarithmic transformations: Reduce skewness in the distribution.
   • Box-Cox
6. Encoding categorical variables:
   • One-Hot Encoding: Convert categorical variables into binary columns -> unordered
   • Label Encoding -> ordered
7. Feature selection.
8. Remove collinearity: It may introduce redundancy into the model. Use correlation matrices and VIF.

Dimensionality Reduction

PCA -> Reduces data size while maintaining the same amount of variance. Component reduction.

LDA (Linear Discriminant Analysis)

Aims to project data into a lower-dimensional space to observe relationships.

• MF and NMF
• T-SNE
• Autoencoder

Division into two sets: training and testing

• 70% training - 30% testing (common)
• 80% training - 20% testing (common)
• 90% training - 10% testing

Cross-Validation (K-fold)

Split the dataset into multiple subsets depending on the K-fold value.

Class Balancing

Standardizing data:

• Oversampling: Increase the number of minority class data by duplicating examples.
• Undersampling: Reduce the number of majority class data by eliminating examples.
• SMOTE: Create synthetic examples of the minority class to balance the classes.

Objectives

• Reduce training time
• Increase accuracy
• Make data compatible with algorithms
• Interpretability