Table of contents
2. Classification of regression models
1. Regression problem
Regression problems contain a causal relationship, and the core is to explore a suitable functional relationship, which represents the mapping relationship from independent variables to dependent variables. For example, given the cardiac function ultrasound measurements of a group of breast cancer patients undergoing chemotherapy, the patient's age, drug use and other information, the regression function can be used to predict the patient's cardiac function ultrasound measurements at a certain point in the future.
2. Classification of regression models
Regression tasks can be divided into linear models and nonlinear models. According to the number of features required by the linear model, it can be divided into a one-element linear model and a two-element linear model.
3. Evaluate the model
There may be multiple regression models for a set of data. Which regression model has a better predictive effect?
Model evaluation with a loss function is required, where two concepts "variance" and "bias" are exposed. Large variance can be understood as model overfitting, that is, the loss value of the model on the training set is small, but the loss value on the test set becomes larger; while large deviation can be understood as model underfitting, that is, the performance of the model on the training set is also low. Poor, indicating that the regression model has not yet learned key features, and may need more training samples and more complex models for training to achieve the best prediction results.
In addition to increasing the data set and optimizing the model, the regularization method can effectively improve the model overfitting problem. Regularization parameters are hyperparameters. Choosing appropriate regularization parameters can help to obtain a model with better performance.
4. Best Model
Generally, choosing the best model depends on the loss function, and the most critical content of the loss function is to use different gradient descent methods to update the weights more quickly and effectively, so that the Loss function converges quickly.