Cross-Validation
The method of measuring the overallization of the model by splitting the data set into different subsets.
Cross verification (cross-validation) is a method used to evaluate overallization performance of a model on the data that does not appear, a limited set of data is more reliable and stable. The most common format is k-fold (k-fold) in cross validation, the data set is divided into random k equal parts; the model is trained once, each time one of the parts is used for the remaining k-1 piece training when separated as a test set. The average and standard deviation of the different performance score obtained at the end of this process offers a much more reliable estimate than a simple train/test separation based on a single data division of the model’s actual performance.
The importance of this method arises, especially in cases where the data set is small: separate the data into a single set of fixed training and testing, depending on which examples are selected for testing, the luck work may give misleading results; cross verification reduces the use of this connection to test every part of the data at least once. Cross verification is also widely used in order to compare the actual overallization performance of different settings in the process of hyperparameter adjustment; special variants that maintain chronological sequence instead of standard k-fold are preferred in sequence data such as time series.
The importance of this method arises, especially in cases where the data set is small: separate the data into a single set of fixed training and testing, depending on which examples are selected for testing, the luck work may give misleading results; cross verification reduces the use of this connection to test every part of the data at least once. Cross verification is also widely used in order to compare the actual overallization performance of different settings in the process of hyperparameter adjustment; special variants that maintain chronological sequence instead of standard k-fold are preferred in sequence data such as time series.