Why is Target Encoding prone to data leakage during the model training process?

Question

Data Science — Hard

Why is Target Encoding prone to data leakage during the model training process?

Accepted Answer

Target encoding replaces categorical labels with the mean of the target variable for that category. If this mean is calculated using the entire dataset before splitting into training and validation sets, information from the target (the 'ground truth') leaks into the training features. This leads to overfitting, as the model 'sees' the target value during training, resulting in artificially high performance metrics.