Detailed explanation of machine learning out of fold prediction | using out of fold prediction oof to evaluate the generalization performance of models and build integrated models

---

One 、 introduction

Model evaluation of machine learning algorithms usually uses Resampling technology , Such as K Crossover verification .

Machine learning models can be used K-Fold Cross validation technology to improve the prediction accuracy of the model . In the process of cross validation , The prediction is performed on the split test set that is not used for model training （ Model training has never seen it ）. These forecasts are called discounted forecasts （out-of-fold predictions）. External prediction plays an important role in machine learning , Sure Improve the generalization performance of the model , as well as Build an integration model .

Summarized below ：

• Out of sample prediction is a kind of out of sample prediction for the data not used in the training model ;
• When predicting invisible data , Discounted forecasts are often used for model evaluation , Prove the generalization performance of the model ;
• Discounted forecasts can be used to build integrated models , It is called stack generalization or stack integration .

The following describes in detail the use of discount forecasting OOF Evaluate the generalization performance of the model and build an integrated model

---

Two 、 What is discount forecast ？

Use resampling techniques such as K-Fold It is a common method to evaluate the performance of machine learning algorithms on data sets .K-Fold The process includes dividing the training data set into K Group , And then use K Each of the group samples is used as a test set , The remaining samples are used as training sets . This means training and evaluating K A different model . This process can be summarized as follows ：

• Randomly scrambling data sets ;
• Divide the dataset into K Group （ Sometimes, it may be necessary to divide by the distribution of labels ）;
• For each unique group ： Use this group as a reserved data for testing , Use the remaining data set as the training set , Fit the model on the training set and evaluate it on the test set , heavy