Fitting and Transforming Data in scikit-learn

A very common job in machine learning is to fit a preprocessing model on a dataset and then transform that same dataset with it. Scikit-learn gives you two ways to write this job. You can call fit and then transform as two separate steps, or you can call fit_transform once and receive the transformed data back directly. Both routes are legitimate, so the practical question is which one to use and why.

Keep this question in mind as you read the code below. If the two approaches give the same answer, the reason to prefer one over the other has to come from somewhere else. The rest of this article works toward that answer.

The snippet below uses StandardScaler from sklearn.preprocessing. The first assignment chains fit and transform together. The second uses fit_transform. The small array X is defined here only so the code can run. The source leaves the definition of X to the reader, so treat this array as a stand-in for any dataset.

Running this code prints two arrays, and the source's point is that they hold the same values. The first line of the chained version does not produce the transformed data directly. It produces the scaler, and the transformed data appears only after the second call. The combined version gets to the transformed data in a single call.

Read the chained line from left to right. Python evaluates scaler.fit(X) first. That call hands back the scaler, so the .transform(X) that follows is called on the returned object. Its output is what gets stored in X_scaled. Following the sequence this way lets you see exactly which steps the code requests, even before you know what each step computes internally.

The source does not describe what the fitting step computes inside StandardScaler, so this article does not go into that. For the purpose here, it is enough to know that fit is the step that works from the data, and transform is the step that applies the result to data.

A useful way to hold the two steps apart is to treat fit as learning and transform as applying. The fit step reads the data and produces a result that the scaler keeps. The transform step takes data and uses that kept result to produce the output. Your data flows through the scaler in the direction shown below.

The source describes fit_transform as the method you use when you want to fit a model on a dataset and then transform it. Every scikit-learn model that has a transform method also has a fit_transform method. So fit_transform is not a special one-off tool for StandardScaler. It is a pattern offered across transformers.

Because the outputs match, the difference is about how the computation is organized. The combined form asks scikit-learn for both steps in one request, and the source says this can often be computed more efficiently. The chained form keeps the two steps visibly separate, which can make the fit and transform stages easier to read in a longer program.

The source explicitly recommends fit_transform when you are transforming the training set. That makes it a sensible default in that case: you fit the preprocessing model on the training data and transform that same training data in one call. Because fit stores what it learns on the scaler, the same scaler can then be used with transform alone on other data.

The test set is handled by transform rather than by a new fit because transform applies the parameters already stored on the scaler. Calling fit on the test set would have the scaler learn from different data, which is not what the training-set guidance describes. Transform on the test set reuses what was learned from the training set.

This means the choice between the two routes is mostly a question of readability and workflow, with efficiency as a possible bonus. When you transform the training set, fit_transform is a sensible default. When a particular transformer or dataset gives no real computational advantage from combining the steps, the chained or separated form loses nothing in correctness, because both produce the same result.