This series is a total of four articles, for heights Field machine Learning Basic study notes. Linear model can get nonlinear model by nonlinear transformation, enhance the knowledge of the model to data, but it leads to a very common problem in machine learning field, overfitting. In order to solve this problem, the rule factor is introduced. In order to solve the selection of the rule factor, the selection of the model and the choice of parameters, the related methods of validation validation are introduced. Machine learning note-Nonlinear Transformation nonlinear\ Transformation Machine Learning Note-Hazard of Overfitting hazard\ of\ Overfitting Machine learning Note-R Egularization Regularization Machine learning Note-Validation Validation What's Overfitting
In the previous article we discussed the linear model linear model Linear\ model plus nonlinear conversion nonlinear transform Nonlinear\ transform can easily produce a nonlinear model to expand our ability to learn, But the disadvantage of doing this is to pay extra model complexity model\ complexity cost. It is this additional model complexity that causes a very easy and difficult problem in machine learning to be over-fitted, and this section first analyzes the causes of the fit and then gives the solution. Gad Generalization
The above is an example of one-dimensional regression analysis. A total of 5 5 data points, X x randomly generated, y y is to bring x x into a two polynomial and then add a little noise noise get. So the best regression curve should be the blue one two times curve. However, because it is not known beforehand, it is possible to use the polynomial of four times to do regression analysis of the 5 5 points to fit. How to use four times polynomial to do regression. 4 4 times the feature conversion + linear regression can get a polynomial of 4 4 times. This will give you a unique 4 4-time polynomial at these 5 5 points.