11. Feature Selection
AI For Trading C6 L1 A09 Feature Selection V3
Understanding Model Optimization
Balancing features and samples is key in model optimization:
More Features Require More Data: Increasing features often necessitates more data. If obtaining real-world samples is challenging, consider techniques like data augmentation or simulation.
Curse of Dimensionality: As features increase, the data space grows exponentially, making data sparse. This can complicate model training and interpretation.
Avoiding Feature Redundancy: Multi-collinearity arises when features are not independent, often detectable through high correlation. Use Variance Inflation Factor (VIF) to diagnose this issue.
Feature Selection Methods:
- Filter Methods: Pre-training techniques using statistical measures like correlation.
- Wrapper Methods: Involve training models with different feature subsets to find optimal performance, though computationally expensive.
- Embedded Methods: Feature selection is part of model construction, such as with Lasso and Ridge Regression.
Dimensionality Reduction: Techniques like PCA aim to project data onto a lower-dimensional space while preserving important information, covered in future lessons.