Machine Learning with Swift

LASSO regression

To fix the problem with irrelevant features, one can replace the L₂ norm with the L₁ norm in a penalty term, and instead of penalizing squares of regression coefficients, penalize their absolute values:

where L₁ norm . This is so-called Least Absolute Shrinkage and Selection Operator (LASSO) regression. Some of the weight coefficients under such a penalty can become exactly zero, which you can think of as a feature selection. If there are several highly correlated features in your dataset, LASSO picks one of them and sets all ...