Table of Contents

WHAT

​

WHY

WHAT

Cross-Validation, or CV for short.

Also called circular estimation, it is a method used to statistically cut a data sample into smaller subsets.

Introduce three nouns:

Training set: learn the sample data set and match the parameters to establish the model.

Validation set: Adjust the parameters of the trained model, and are also used to determine the network structure or parameters that control the complexity of the model.

Test Set: Test the model.

Three important indicators:

Bias: Accuracy.The degree of deviation between the expected prediction of the learned model and the actual result (difference between the average predicted value and the actual value), which is used to describe the fit degree of the algorithm itself.

Variance: Stability.The performance change (expectation of the square of the difference between the predicted value and the average predicted value) during training with different training sets of the same size is used to characterize the impact of data perturbations.

Error: The accuracy of the entire model.

Note:

1, Error=Bias^2+Variance+Noise

2. Bias and Variance are often not compatible.Both are low, which is an ideal state (see the figure below), but if you want to reduce the Bias, the Variance will increase to a certain extent, and vice versa.

The root cause: We prefer to use a limited sample data set to estimate and predict an infinite real data set.When we continuously improve the accuracy of the model (Bias is reduced), overfitting will occur, the generalization ability of the model will be reduced, the performance of the model in real data will be reduced, and the uncertainty of the model will be increased (Variance increases).Conversely, if more restrictions are added in the process of learning the model, the Variance of the model can be reduced (Variance reduction) and the stability of the model can be improved, but the Bias of the model will be increased (Bias increased).

Summary: Overfitting has high bias and underfitting has high variance.

So, how to avoid these two extreme cases???

(1) Avoid underfitting: find better features (representative), use more features (increase the dimension of the input vector).

(2) Avoid overfitting: increase the data set (reduce the proportion of noise), reduce data features (reduce data dimension), regularization method (add a regular term to the objective function or cost function), Cross-validation method (the key part of this post)

Three CV methods: Hold-out Method, K-fold Cross Validation, Leave-One-Out Cross Validation

Details about K-fold Cross Validation here

1. Divide the original data into k groups (usually equally divided),

Each subset is used as a validation set, and the remaining k-1 sets of subset data are used as training set

Get k models

2. Use the average of the classification accuracy of the final validation set of the k models as the performance index of the classifier under this k-CV

3. Evaluate the effect of the k models and pick the best hyperparameters (hyperparameters are parameters that set values ​​before starting the learning process, not parameter data obtained through training.).

4. Use the optimal hyperparameters, and then retrain the model with all the k data as the training set to obtain the final model.

WHY

1. Cross-validation is used to evaluate the prediction performance of the model, especially the performance of the trained model on new data, which can reduce overfitting to a certain extent.

2. Obtain as much effective information as possible from limited data.

3. A convenient technique to measure model performance using only the training set, instead of using the test set after modeling.