The fit Method in scikit-learn

Every scikit-learn estimator goes through a two-part life. First you fit it on data, then you put it to work. Fitting is the step where the estimator learns from data, and it is easy to think of it as the finish line. It is not. Fitting only learns from data. By itself it does not tell you anything about new data.

The question that follows fitting is what you want back from the model. The answer splits into two broad families. One family returns a new output for each input. The other returns a reshaped version of the inputs. A third method, score, serves a separate purpose: evaluating a supervised model. The methods in this article are predict, transform, and score.

Fitting is the step that gives the other methods something to work with. During fit, the estimator takes in training data and learns from it. The learned result is then available to the methods that come later. Those later calls do not relearn anything from scratch. They reuse what fit learned, applied to whatever inputs you pass in.

The diagram shows the order of calls. Fit comes first and does the learning. The later calls can happen in any order you need, and each one applies the learned state to inputs you provide. Once you see fit as the setup step, the other methods read as different ways of using the same fitted object.

Here X_test is the set of inputs you want to predict for, and y_pred is the output the estimator produces for them. The method is used for classification, regression, and clustering. In each case the output is a new value or label attached to each input. That is why predict is the method you reach for when the question is what the model thinks each input corresponds to.

The transform method answers a different question. Instead of producing an output such as y, it creates a new representation of the input data X. The call has the same shape as predict, but the result is a different kind of object.

The result X_new is still a representation of the same inputs, but it is no longer the original data. The typical uses of this method are preprocessing, dimensionality reduction, feature extraction, and feature selection. Each of these changes how the inputs are described, not what the model guesses about them. This is the key contrast with predict. A transform call hands back reshaped inputs, while a predict call hands back a prediction.

A helpful way to pick a method is to ask what the result is for. Predict results are typically consumed as answers. Transform results are typically passed onward, either to another step in a pipeline or to a model that will consume the new representation. Score sits apart from both, because its purpose is evaluation rather than output or representation.

Read the flowchart from the top. If the goal is a prediction such as what class or value each input belongs to, call predict. If the goal is to change how the inputs are described before a later step, call transform. If the goal is to judge how well a supervised model is doing, call score.

The third method is score. Its purpose is evaluation rather than output or representation. All supervised models have a score method that allows an evaluation of the model. It takes two inputs: the test data and the test labels.

Because score needs y_test, it can only be used when you have labels for the inputs you are evaluating. This is why the method is tied specifically to supervised models. The data flows in two directions at once: the inputs go through the fitted model, and the model's results are compared against the true labels to produce one number that summarizes how well it did.

The three methods share a similar call pattern, which makes them easy to confuse. The following mistakes come up often when a learner first applies them.