Building and Evaluating a k-Nearest Neighbors Model

A trained k-Nearest Neighbors model is only useful if we can say how often it is right. Instead of guessing, we check its predictions against answers we already know. This article follows one short piece of Python that does exactly that: it asks a model for predictions, then reports a single number that summarizes how well those predictions matched reality. The goal is to understand what that number means, where it comes from, and what it does not promise.

The flowers used to evaluate the model are the test set that was set aside earlier. Those flowers were not used to build the model. That separation is what makes the evaluation meaningful: the model is being asked about flowers it has not already seen, and we know the correct species for each one because the labels were held back from the model during building.

Read the diagram from top to bottom. The full set of flowers with known species is split once. The training part is what the model stores; the test part is kept aside. The model is built from training examples only, then it is asked about the test inputs. Its predictions meet the known labels only at the final comparison. If the test examples had leaked into the building step, the score would describe flowers the model had already seen, and it would not measure performance on new data.

The evaluation has two inputs that must line up. The first is the model's prediction for each test flower. The second is the known label for that same flower. Each prediction is compared against its own label, and the results are summarized. If predictions and labels were misaligned, the number would be meaningless even though the code ran without error.

The index map shows the comparison as it actually happens. Each position holds one prediction and one label for the same flower. Position 2 is the only mismatch, so it is the only one that does not count toward the score. The comparison step does not look at species names as a group; it checks each position on its own, and only then is the result summarized.

The source shows two ways to obtain the test set accuracy. The first builds the comparison by hand: it checks equality between the predicted labels and the known labels, then takes the mean of the resulting True and False values. The second calls the score method of the trained model object, which performs the same test set accuracy computation internally. Both produce 0.97 in the source, so they agree on this model and this test set.

The source interprets a test set accuracy of about 0.97 as meaning the model made the right prediction for 97 percent of the irises in the test set. That statement is a direct description of the test set. Under some mathematical assumptions, the same figure is taken as an expectation that the model will be correct about 97 percent of the time on new irises. For the hobby botanist in the source, that level of accuracy suggests the model may be trustworthy enough to use.

The Nearest Neighbors model is one of the models you can choose from when you are deciding which machine learning approach fits a problem. Its name describes its core idea: to make a decision about a new data point, the model looks at the existing data points that are nearest to it. The name refers to that act of looking for neighbors, meaning the stored examples that sit closest to the new point. The letter k in k-Nearest Neighbors is the number of such neighbors that are consulted; the lookup in the next section uses only the single nearest one, which is the simplest case.

The example below is an original illustration, not taken from the source. It uses a tiny dataset of single numbers, each paired with a label, and finds the stored example closest to a new query value. Real problems usually involve many measurements per example, but the logic of comparing and picking the closest one is the same.

Work through the lookup by hand to check the code. The query 3.4 is 2.4 away from 1.0, 0.9 away from 2.5, 0.6 away from 4.0, and 3.6 away from 7.0. The smallest distance is 0.6, which belongs to the stored example 4.0 with label A. That stored example is the nearest neighbor, and its label is the answer this lookup gives.

The source identifies small datasets as the setting where this model is a good fit. When a dataset has relatively few examples, a model that works directly from those examples is practical to try, and it gives you a clear starting point before you commit to more complicated approaches. Keep in mind that the source states this suitability without giving a size threshold, so you should judge what counts as small for your own problem.

A baseline is a simple reference model. Its job is not necessarily to be the final choice, but to give you a point of comparison. If a more complex model cannot clearly beat a simple baseline, that is useful information about the problem and about the extra complexity. The source describes the Nearest Neighbors model as good as a baseline, which means it is a sensible first model to build and measure before you move on to others.

The idea behind the name is simple to describe to someone who has never studied machine learning. A new point is compared with the stored examples, and the ones closest to it are the ones that matter. Because the reasoning can be stated in one sentence, the model is easy to explain to colleagues, stakeholders, or anyone who needs to understand why a prediction was made. This explainability is one of the reasons the source recommends it as an early model.

When a reported score is not what you expected, the flow from the earlier sections gives you a checklist. Work through the stages in order. Only after those checks does it make sense to question the model.