Técnicas de regularización en el marco del aprendizaje de máquina: regresiones ridge y lasso

Abstract

Linear regression is one of the most widely used machine learning methods today. However, the standard ordinary least squares method makes several assumptions about the data that are often not true in real-life data sets. This can cause numerous problems when the model is fitted using least squares. One of the most common problems is that the model fits the data too much, this happens when the estimator is unbiased, but has high variability. The Ridge and Lasso regressions are two regularization techniques used to create a better and more accurate model. This work explains how high variability occurs in the least squares estimator. An example is included with a real-life data set and these methods are compared with the least squares estimator to infer the benefits and drawbacks of each method.