Target Encoding

Categorical features describe groups such as depots, suppliers or product types. One-hot encoding creates a separate indicator for each category. With many categories, target encoding offers another representation: replace each category with an outcome summary, such as a mean delay for regression or a positive-outcome rate for binary classification. The useful signal is also the source of risk. Imagine a category that appears once. Its unsmoothed category mean is that row's outcome. If this value becomes an input for predicting the same row, the model is being given the answer. Strong training performance in this setup can disappear on new data. Cross-fitting helps construct safer training representations. Split the training data into folds, calculate category summaries using the other folds, and encode the held-out fold with those summaries. Repeat until each training row has an encoding built without its own fold's labels. The downstream model learns from these representations. At evaluation or prediction time, apply mappings learned from the permitted training data. Smoothing reduces the influence of categories with few observations by pulling their summaries toward a global training mean. A rare depot should not receive an extreme encoding solely because of one unusual delivery. The amount of smoothing is a modeling choice that needs validation. Scikit-learn's TargetEncoder documents an important distinction: fit_transform uses internal cross-fitting, while fitting and then transforming the same training data does not provide that equivalent protection. Do not assume every library implements the same behavior. Keep the encoder within the evaluation pipeline, define how unseen categories are handled, and adapt the splitting strategy when time order or repeated entities make ordinary random folds inappropriate.