L1 regularization
Penalises the absolute size of the weights instead of the square. The constant gradient near zero drives weights exactly to zero rather than merely close, which is what makes it a feature selector.
- The regularisation strength.
- The absolute value of one weight.