Regularization and GeometryMinh PhamBlockedUnblockFollowFollowingMay 11I.
Bias-Variance TradeoffWhen we perform statistical modeling, the goal is not to select a model that fits all the training data points and obtain the smallest error on the training data.
The objective is to give the model the ability to generalize well on new and unseen data.
Bias and variance contributing to generalization error As more and more parameters are added to a model, the complexity of the model increases, i.
, it can fit more noise in training data and leads to overfitting.
This means that we are increasing the variance and decreasing the bias of the model.
From the figure above, if we keep raising the model’s complexity, the generalization error will eventually pass the optimal spot and continue to increase.
To conclude, if our model complexity exceeds the optimal point, we risk falling into the overfitting zone; while if the complexity falls short of the optimal point, we are in the underfitting zone.
Regularization is a technique that helps prevent the statistical model from falling into the overfitting zone.
It does so by hindering the complexity of the model.
RegularizationLinear RegressionThis is the equation for a simple linear regression.
Linear RegressionOur goal is to find the set of beta to minimize the residual sum of squares (RSS).
Residual Sum of SquaresL2 RegularizationA regression model that utilizes L2 regularization is also called Ridge Regression.
In terms of Linear Regression, instead of just minimizing the RSS.
We add another piece to the game by having a constraint for the betas.
L1 Regularization“s” can be understood as a budget for the betas.
If you want to make one of the first beta big, the second beta has to be small and vice versa.
This becomes a competition and as we have more beta’s, the unimportant ones will be forced to be small.
Due to the square term, the RSS equation actually gives us the geometry of an ellipse.
The black point is the set of beta’s using the normal equation, i.
, without regularization.
The red circle contains all the sets of beta’s that we can “afford”.
The red point is the final beta’s for Ridge Regression, which is closest to the black point while in our budget.
The red point will always be on the boundary of the constraint.
The only time that you attain a red point inside the constraint is when the solution for the normal equation, i.
, the solution for Linear Regression without regularization, already satisfies the constraint.
L1 RegularizationA regression model that utilizes L1 regularization is also called Lasso (Least Absolute Shrinkage and Selection Operator) Regression.
The goal of Lasso is similar to Ridge with the exception that the constraint becomes:L2 RegularizationLasso also provides us with a beautiful geometry that comes with unique properties.
The set of beta’s that we can “afford” with L2 regularization lies within a diamond.
The red point is at the corner of the diamond, which sets one of the betas to 0.
What if the red point is on the edge of the diamond instead, i.
, the ellipse touches the diamond on the edge? Hence, neither of the beta’s will be 0.
However, the diamond in 2-D space is a special case where neither of the beta’s becomes 0 if the red point is on the edge.
Let us consider the case where we have 3 beta’s.
The geometry for the constraint becomes:Lasso with 3 beta’sThe figure above consists of 6 vertexes and 8 edges.
If the red point lies on an edge, one beta will be set to 0.
If it lies on a vertex, two beta’s will be set to 0.
As the dimension increases, the number of vertexes and edges increases as well, making it more likely that the ellipse will be in contact with the diamond on one of those places.
That being said, Lasso tends to work better in higher dimension.
Differences between L1 and L2 RegularizationRidge regression is an extension for linear regression that enforces the beta’s coefficients to be small, reducing the impact of irrelevant features.
This way the statistical model will not fit all the noise in the training data and fall into the overfitting zone.
Lasso regression bring some unique properties to the table because of its beautiful geometry.
Some of the beta’s will be set to 0, giving us a sparse output.
We can also use Lasso for feature selection.
While feature selection techniques such as Best Subset, Forward Stepwise or Backward Stepwise may be time inefficient, Lasso will converge faster to the final solution.
ConclusionA statistical model that with high complexity may be prone to overfitting.
In this article, I introduced two regularization techniques to discourage the model from fitting all the noise in the training data.
Moreover, I explained their properties and differences with the help of geometry.
That is the end of my article! Have a wonderful day folks :)Images: Daniel Saunders, The Bias-Variance Tradeoff (2017).. More details