Data Leakage
Due to the fact that the information of the test data is accidentally mixed in the educational process, the performance of the model looks misleading.
Data leakage (data detox) is the condition that normally the model should not be accessed during the training, the knowledge of the test or real world use is much higher and misleading as the result of the performance of the model, accidental training process. This is often fine, the difference is caused by difficult mistakes: for example, scaling the data before splitting into training and test clusters can lead to leakage of the statistics in the test set into the educational process; or in a hospital data set, it can lead to the appearance of multiple records belonging to the same patient, as "heavy" the model actually does not learn. Time series data is also an example of classic leakage to predict the history of the future information accidentally.
The danger of data leakage is that a model that gives impressive results in the laboratory environment leads to unexpectedly poor performance when it comes to the actual production environment because the model has learned based on "taking" information that will never be accessed in the real world. Before pretreatment of all kinds of data partition processing to avoid this problem, it is one of the most basic and most often neglected disciplines of machine learning engineering, in accordance with the structure of careful and data set.
The danger of data leakage is that a model that gives impressive results in the laboratory environment leads to unexpectedly poor performance when it comes to the actual production environment because the model has learned based on "taking" information that will never be accessed in the real world. Before pretreatment of all kinds of data partition processing to avoid this problem, it is one of the most basic and most often neglected disciplines of machine learning engineering, in accordance with the structure of careful and data set.