Training a k-Nearest Neighbors Classifier on the Iris Dataset

So far, the model you built learned from flowers whose species were already recorded. The real payoff comes when you meet a flower whose species nobody has written down. Suppose you measure a flower and get a sepal length of 5 cm, a sepal width of 2.9 cm, a petal length of 1 cm, and a petal width of 0.2 cm. You want the trained model to name its species. This article follows that measurement from your notebook into a prediction.

Keep one question in mind as you read: at which step does the flower stop being a list of numbers and start being a species name? Following that transformation is the goal of this article, and it is also the place where most beginners get confused.

New data for prediction must be packaged as a two-dimensional NumPy array with shape (n_samples, n_features). The first axis counts samples, which are the flowers you are asking about. The second axis counts features, which are the measurements taken on each flower. A single flower is therefore stored as one row.

The most important detail is the doubled pair of square brackets. A single flower could be written as a flat list of four numbers, but scikit-learn does not accept that shape for prediction. It always expects a two-dimensional array. Even when you have only one flower, it still needs a row, so the array has to be nested one level deeper than the list of numbers.

Read the structure from the outside in. The outer brackets make the array two-dimensional by listing rows. The inner brackets form the single row for this flower. The four numbers fill that row in the same order the features were measured. Changing the order of the numbers would change which measurement the model treats as which feature, so the order matters as much as the brackets.

Once the array is in the right shape, you call the predict method on the trained knn object and pass it the new array. The method uses the model you already built to decide which class the new flower belongs to, and it returns a predicted class label. In the source example that label is stored in a variable named prediction. The printed output shows it inside square brackets, which tells you the result is an array with one entry for each sample you passed in.

Run the code and compare each printed line with your own guess. The first line only checks the array you built. The lines after the predict call depend on the trained model, and they show the two forms of the answer: the raw label the model returns, and the species name that the dataset attaches to that label.

Follow the flower through the whole pipeline. The four measurements become a two-dimensional array. The predict method turns that array into a label. The dataset's target names turn the label into a word a person can read. Each step is a handoff, and each handoff changes what the data represents.

Notice that the model never produces a species name directly. It produces a number, and the name is recovered afterward from the dataset. This separation is why the question of where the flower stops being numbers and becomes a name has a precise answer: it happens at the target_names lookup.

The class label is a number. Numbers are convenient for the model but not for people. The target_names attribute of iris_dataset solves this: it holds the species names, and the predicted label can be used to look up the matching entry. The source example performs this lookup by indexing iris_dataset['target_names'] with the prediction. The label works as a position: it selects one entry from the stored list of names.

Because the prediction comes back wrapped in an array, the lookup also returns an array-style result. That is why the printed name appears inside brackets in the source output. The value you read as the species name is the single element inside that list. If you need a plain string, you would take that one element out of the array. The source example keeps the output in the array form shown above.

The model has now produced an answer, and it is tempting to treat that answer as settled. The right question is how we know whether we can trust it. For this flower, nobody has told us the correct species. That gap is the reason the model was built in the first place. A prediction on unlabeled data can be useful, but its correctness cannot be checked against a known answer at the moment you make it.

This does not mean predictions are worthless. It means that judging how far to rely on them is a separate step from making them. Making a prediction is one step. Deciding whether it can be trusted is another, and the earlier evaluation of the model is what gives you grounds to make that decision.