An online advertising company is developing a linear model to predict the bid price of advertisements in real time with low-latency predictions. A data scientist has trained the linear model by…

L1 regularization (Lasso regularization) adds a penalty to the loss function that is proportional to the absolute value of the model's weights. This encourages sparsity by driving the weights of less important features exactly to zero, effectively performing automatic feature selection and reducing the number of features while preventing overfitting.

L2 regularization (Ridge regularization) adds a penalty proportional to the square of the weights. While it helps prevent overfitting by shrinking weights, it rarely drives them completely to zero, so it does not primarily reduce the number of features.

Dropout regularization is primarily used in neural networks to prevent overfitting by randomly dropping units, but it is not typically applied directly to linear models to reduce the number of features.

Retraining the model with more data can help reduce overfitting, but the question specifically asks to reduce the number of features, which L1 regularization directly addresses more effectively than simply adding more data.