09. Feature Crosses

Jump to bottom

Antonio Erdeljac edited this page Feb 26, 2019 · 1 revision

Feature Crosses

Topic: Feature Crosses

Course: GMLC

Date: 20 February 2019

Professor: Not specified

Resources

Key Points

Feature cross
- Synthetic feature (Feature that is created by combining 2 features) that multiplies two or more input features thus encoding nonlinearity in feature space
- X3 = x1 * x2
- Y = w1x1 + w2x2 + w3x3
- Linear algorithm learns weight for w3 normally as for others
Types of feature crosses
- Two features [A x B]
- Multiple features [A x B x C x D x E]
- Squaring a single feature [A x A]
Crossing one-hot vectors
- One hot vector - Feature vector created using a single categorical feature [1,0,0,0…]
- Example using binned longitudes & latitudes
  - Binned_longitude = [0, 0, 0, 1]
  - Binned_latitude = [0, 1, 0, 0]
  - binned_longitude x binned_longitude - crossed one hot vector
  - Creating a crossed one hot vector we can vastly improve predicting ability of a model
- Example using dog owner satisfaction:
  - [behaviour_type x time_of_day]

Check your understanding

Explain feature crossing & its purpose

Summary of Notes

Feature cross Is a synthetic feature used to improve model’s learning by combining 2 or more features and solve non-linear problems by encoding (binning) them