Descripción
|
|
---|---|
In this talk we will present two recent contributions within the field of regularization. Regularization techniques provide estimates for the linear regression coefficients solving the problems encountered in the "few samples and many variables" setting. The main idea is to shrink the coefficients to zero by imposing a penalty on their size. We will firstly review the main regularization techniques. We will then propose a method for the structure learning of Gaussian Bayesian networks. The search in an equivalence class search space is combined with regularization techniques, promoting a sparse network learning. Finally, a new regularized logistic regression method based on the evolution of the regression coefficients using estimation of distribution algorithms is presented. The main novelty is that it avoids the determination of the regularization term. The chosen simulation method of new coefficients at each step of the evolutionary process guarantees their shrinkage as an intrinsic regularization. | |
Internacional
|
Si |
Nombre congreso
|
Seminario para el Department of Computer Science |
Entidad organizadora
|
Faculties of Engineering, Science and Medicine |
Nacionalidad Entidad
|
DINAMARCA |
Lugar/Ciudad de impartición
|
Aalborg |
Fecha inicio
|
29/05/2009 |
Fecha fin
|
29/05/2009 |