NumPy Arrays for Machine Learning

Before a machine learning model can learn anything, the data has to arrive in a form the model can read. Many introductory examples in scikit-learn use the Iris dataset, and the way that dataset is organized is a good template for thinking about almost any labeled dataset. Instead of one tangled object, the Iris data is packaged as two separate pieces that work together.

The question for this article is simple: what are those two pieces, and how do they relate to each other? By the end, you should be able to look at the two arrays, say what each one contains, and follow a single flower from one array to the other.

The Iris dataset consists of two NumPy arrays. The first holds the data, which scikit-learn refers to as X. The second holds the correct or desired outputs, which is called y. The division is deliberate: the inputs the model looks at are kept apart from the answers it is supposed to predict.

Keeping the two arrays separate makes the roles clear. X is the question the model looks at, and y is the answer it is supposed to predict. When you read code that works with this dataset, the name tells you the role: anything called X is input, and anything called y is the target.

X is a two-dimensional array of features, with one row per data point and one column per feature. Reading the shape of X is a useful habit. If X has $n$ rows and $d$ columns, then it has $n$ data points and $d$ features per data point.

Two dimensions are needed because the data is a table. One axis selects which sample you are looking at, and the other selects which feature of that sample. The rows run across samples and the columns run across measurements taken on each sample. So reading across a row gives you every measurement for one flower, while reading down a column gives you one measurement for every flower.

The example below builds a tiny version of the same structure by hand, so you can check every number. The first array X holds three samples with four features each, and the array y holds one integer label per sample.

The easiest way to see how X and y fit together is to follow one data point through both arrays. The row of X for that data point gives its features, and the entry of y at the same position gives its class label. The position is the link between the two arrays.

The diagram shows the pairing for the first three samples. Each row of X is linked to the entry of y at the same position. When you pick out sample 2, you take row 2 from X for its features and entry 2 from y for its label, and both belong to one flower. The pairing depends on position, so the arrays stay aligned only as long as their order is kept.

Because y is one-dimensional, it has no columns. It is a single list of answers, one per data point. The integers 0, 1, and 2 are class labels, which means they name categories rather than measuring quantities. They mark which of three categories a sample belongs to.