Your First Machine Learning Application: Iris Classification

Imagine you copy a working example from a tutorial onto your own computer, run it, and get an error about a module that the tutorial clearly imports. Nothing in the code itself is wrong. The difference is in the library. Machine learning code depends on a specific library, and that library changes over time. A program can be perfectly written and still fail if the library installed on your machine does not provide the pieces the program expects.

This article is about that dependency. The library in question is scikit-learn, the machine learning toolkit used throughout the examples that follow. Before you write a single line of modeling code, you need to know which version of scikit-learn your code expects, and whether the version on your machine is recent enough to supply it.

Before running the course examples, you can ask Python which scikit-learn version it is actually using. The check imports the library and prints its version string. You then compare the printed value against the 0.18 minimum. Treat this as the first step of every session: a version check costs seconds, while a confusing import error can cost much longer to diagnose.

The course assumes that you have scikit-learn version 0.18 or later. Notice what this requirement does and does not say. It is a floor, not an exact match. The course does not ask you to install one particular release. It asks that your version be at least as recent as the one used when the material was written.

The reason 0.18 is the threshold is a specific feature: the model_selection module was added in that version. Many of the tools you will use to split data and evaluate models live inside this module, so the course code reaches for it directly. If your installed version predates it, the import statements that refer to that module will not match what your library provides.

Now follow a learner whose installed version is older than 0.18. The learner keeps the course code exactly as written. The first point of trouble is the import that pulls tools from model_selection. Because that module did not exist yet in the older release, the import no longer matches the library. The learner is not expected to rewrite the modeling logic. The source's advice is narrower: the imports from this module must be adjusted.

Version compatibility is about running the right code, but the code itself solves a particular kind of problem. The course's first application comes from a hobby botanist who has measured irises. Each flower has measurements of petal length and width and sepal length and width, all in centimeters. Some of these irises have already been identified by an expert as belonging to one of three species: setosa, versicolor, or virginica. The botanist can be certain of those species, and she assumes these are the only species she will meet in the wild.

The goal is a model that learns from the measurements of irises with known species, so it can predict the species of a new iris. Because the correct answer is known for every measured flower in the training set, this is called a supervised learning problem. The answers supervise the learning: the model is shown both the measurements and the correct species.

The labeled dataset works because each row of measurements is paired with a label at the same position. Sample 0 is the first flower, and its four measurements sit in the first row of the feature data. The label for sample 0 sits at position 0 of the label array. Position i in the labels array therefore refers to the same flower as row i of the feature data. Nothing in the data links them except that shared position, which is why the order of rows must never be shuffled out of step with the labels.