Data Splitting for Model Evaluation

A very common first reaction to MinMaxScaler is to assume that scaling puts every feature into the same tidy box. You fit the scaler, you transform the data, and you expect every number in the result to sit between 0 and 1. For the training data, that expectation is correct. For the test data, it is often wrong. The mismatch is a useful place to learn how scaling actually works, and this article follows one workflow from start to finish: load a dataset, split it, scale it, and check the results.

Preprocessing methods such as scalers are usually applied before a supervised algorithm is applied. The model, however, must still be evaluated on data it has not seen. That is why the dataset is split into a training set and a test set before any scaling happens. The test set is needed to evaluate the supervised model afterward, so it has to remain a separate collection of rows that the model and the preprocessing steps are not allowed to treat as already known.

The running example uses the cancer dataset from an earlier chapter. The scenario is to train a kernel SVM, which is a supervised model, and to preprocess its inputs with MinMaxScaler. The dataset has 569 data points, and each one is described by 30 measurements. The split sends 426 samples to the training set and 143 samples to the test set. The shapes printed by the code confirm that the feature count is still 30 in both sets; only the number of rows is divided.

Once the split is done, scaling follows a fixed sequence. The preprocessing class is imported and then instantiated, just as the supervised models from earlier were created. The instance is then fitted, transformed on the training data, and transformed on the test data. Each step has a specific job, and the order matters.

Fitting is the step that learns from data. For MinMaxScaler, the fit method computes the minimum and maximum value of each feature on the training set. Unlike the classifiers and regressors used earlier, the scaler is given only the training features when fit is called. The target values y_train are not passed in and are not used. The scaler receives only the features, not the labels.

Fitting only records the numbers. Applying the change is the job of the transform method. In scikit-learn, transform is the method used whenever a preprocessing object returns a new representation of the data. Calling it on the training set produces the scaled training data. The shape stays the same, because the features are shifted and scaled rather than removed.

In words, transform subtracts the training set minimum and divides by the training set range. The minimum and the range are learned once, during fit, from the training data. The same two numbers are then reused for every later call to transform, including the call on the test set.

The cancer dataset is too large to inspect by hand, so this example uses one feature with a handful of values. Trace the state of the scaler at each step. The arithmetic is simple enough to check on paper before running any code.

The printed minimum and maximum for the test set are not 0 and 1 in the cancer workflow, and some scaled features fall outside the 0 to 1 range. The source explains this directly: MinMaxScaler, like the other scalers, applies exactly the same transformation to the training and test sets. The transform subtracts the training minimum and divides by the training range. The test set's own minimum and range might be different, so its scaled values can land outside the training box.