Wherein Penalised Regression Is Presented as a Remedy for Ill‑conditioned Inverse Problems, and L2 (Ridge) Penalties Are Noted for Yielding Tractable Information‑criteria and a Bayesian Prior Interpretation.
Regression estimation with penalties on the model parameters. I am especially interested when the penalties are sparsifying penalties, and I have more notes on sparse regression.
Here I consider general penalties: ridge etc. At least in principle — I have no active projects using penalties without sparsifying them at the moment.
In nonparametric statistics, we might estimate many parameters simultaneously, constraining them in some clever way, which usually boils down to something we can interpret as a “penalty” on the parameters.
“Penalization” has a genealogy unknown to me but is probably the least abstruse for common, general usage.
The “regularisation” nomenclature claims descent from Tikhonov, (e.g. Tikhonov and Glasko (1965)) who wanted to solve ill-conditioned integral and differential equations, which is slightly more general.
In statistics, the term “shrinkage” is used for very nearly the same thing.
“Smoothing” seems to be common in the spline and kernel estimate communities of (Silverman 1982, 1984; Wahba 1990) et al, who usually actually want to smooth curves. When we say “smoothing” we usually mean that you can express your predictions as a “linear smoother”/hat matrix, which has certain nice properties in generalised cross validation.
“Smoothing” is not a great general term since penalisation does not necessarily cause “smoothness” from any particular perspective — for example, some penalties cause the coefficients to become sparse and therefore, it promotes non-smooth vectors. Often the thing that becomes smooth is not obvious.
Regardless, what these problems share in common is that we wish to solve an ill-conditioned inverse problem, so we tame it by adding a penalty to solutions we feel one should be reluctant to accept.
🚧TODO🚧 specifics
1 Connection to Bayesian priors
Famously, a penalty can have an interpretation as a Bayesian prior on the solution space. It is a fun exercise, for example, to “rediscover” the lasso regression as a typical linear regression but with the plus prize for the coefficients. In that case, the maximum a posteriori estimate given those bays rise and the lasso solution coincide. If you want to know the full posterior, you have to do a lot more work. But the connection is suggestive nonetheless.
A related and useful connection is the interpretation of covariance kernels as prize producing smoothness in solutions. A very elegant introduction to these is given in Miller, Glennie, and Seaton (2020).
What should we regularize to attain specific kinds of solutions?
Here’s one thing I saw recently:
Venkat Chandrasekaran, Learning Semidefinite Regularizers via Matrix Factorization
Abstract: Regularization techniques are widely employed in the solution of inverse problems in data analysis and scientific computing due to their effectiveness in addressing difficulties due to ill-posedness. In their most common manifestation, these methods take the form of penalty functions added to the objective in optimization-based approaches for solving inverse problems. The purpose of the penalty function is to induce a desired structure in the solution, and these functions are specified based on prior domain-specific expertise. We consider the problem of learning suitable regularization functions from data in settings in which prior domain knowledge is not directly available. Previous work under the title of ‘dictionary learning’ or ‘sparse coding’ may be viewed as learning a polyhedral regularizer from data. We describe generalizations of these methods to learn semidefinite regularizers by computing structured factorizations of data matrices. Our algorithmic approach for computing these factorizations combines recent techniques for rank minimization problems along with operator analogs of Sinkhorn scaling. The regularizers obtained using our framework can be employed effectively in semidefinite programming relaxations for solving inverse problems. (Joint work with Yong Sheng Soh)
Kloft, Rückert, and Bartlett. 2010. “A Unifying View of Multiple Kernel Learning.” In Machine Learning and Knowledge Discovery in Databases. Lecture Notes in Computer Science.