Before moving on to linear classifiers, it makes sense to also add the following additional linear regression algorithms to your toolset:
- Elastic-net uses a mixture of L1 and L2 regularization techniques, where l1_ratio controls the mix of the two. This is useful in cases when you want to learn a sparse model where few of the weights are non-zero (as in lasso) while keeping the benefits of ridge regularization.
- Random Sample Consensus(RANSAC) is useful when your data has outliers. It tries to separate the outliers from the inlier samples. Then, it fits the model on the inliers only.
- Least-Angle Regression (LARS) is useful when dealing with high-dimensional data—that is, when there is a significant number of features compared to the number of samples. You may want to try it with the polynomial features example we saw earlier and see how it performs there.
Let's move...