exploratory-data-analysis.md

Feature Engineering

• label encoding
  • cardinal to ordinal bias
• one-hot-encoding
  • frequency-based encoding
• target encoding
  • integrating label information to itself, easy to overfit
    • separate part of data solely used for encoding
• word2vec
• too many features
  • frequency-based encoding, target-encoding
  • hashing or clustering to combine features
• too sparse features

Missing Value Handling

• Common Methods - Analyze Impact
  • Deletion
  • Imputation
• Categorical Features
  • Make NA as one category
  • Logistic Regression (Model)
• Continuous Features
  • Mean, Median, Mode, etc
  • Linear Regression
    • iterative
    • introduces some degree or correlation among features, but can me remediated by separating into sub-groups

Feature Engineering Basics

• Imbalanced Feature
  • Upsampling
  • Downsampling
• Spare Feature
  • Tree-based model are bad with sparse features
  • Use **Embeddings **to transfer higher dimensional data to lower dimensions
    • Tree based model + Logistic Regression/Neural Network