Parent: data-representation

Source: google-ml-course

Scaling/normalising feature values

• No benefit if only one feature in the feature set
• Can be beneficial for multifeature set
  • Helps the gradient descent to converge more quickly
  • Helps avoid numbers becoming NaN
  • Makes the model spend the same time learning for each feature vector. If scaling is not done, more learning is done for features which have a bigger range of data.
• Scaling methods
  • Map [min, max] to [-1, +1]
  • Use Z scores
    $$ Z = \frac{x - \mu}{\sigma} $$
    which results in most values lying within [-3, +3]
  • Log scaling for data with very big difference between min and max value (might still not be very ideal)
• Clipping/capping at a max value (this doesn’t mean all values exceeding the max are thrown out, instead the outlier values get mapped to the max value)

Original	Log	Clipped