Loading Built-In Datasets in scikit-learn

When you learn a machine learning algorithm, you need data to test it on. The datasets used for this purpose come in two broad kinds. Some are small and synthetic, which means they are made up. Others are large and come from real-world applications. Both kinds have a role, and understanding why each one exists makes the rest of the algorithm material easier to follow.

A real-world dataset works the other way around. It is collected from an application that exists outside the textbook, so its structure is whatever the real process produced. Nobody arranged its features to make one algorithm look clear. The Wisconsin Breast Cancer dataset, often called cancer for short, is an example. It records clinical measurements of breast cancer tumors, and each tumor is labeled as benign, meaning harmless, or malignant, meaning cancerous. The task is to learn to predict whether a tumor is malignant from measurements of the tissue.

Neither kind replaces the other. The synthetic sets make a specific behavior easy to see, and the real-world sets show whether that behavior appears in practice.

The first dataset this course uses is the forge dataset. It is a synthetic two-class classification dataset with two features. The data is split into two parts. X holds the data points, and y holds the class labels. The printed shape is the fastest way to check what a dataset contains.

Because the labels in y are kept separately, with one label per point, they line up with the rows of X. Row 0 of X and the first entry of y describe the same data point. Reading the shape therefore tells you the size of both parts at once: 26 points, each with 2 feature values, and 26 labels.

Printing a shape tells you how big the dataset is, but not what it looks like. A scatter plot fills that gap. In the forge plot, the first feature sits on the horizontal axis and the second feature sits on the vertical axis. Each data point becomes one dot, placed at the coordinates given by its two feature values.

The color and shape of each dot show the point's class, so you can see at a glance which points belong to Class 0 and which belong to Class 1. This is the main reason scatter plots are useful for classification data. They let you check whether the classes separate, overlap, or are mixed together. Patterns in the dots, not in the printed numbers, reveal how the features relate to the classes.

The forge dataset answers a classification question: which class does a point belong to? For regression algorithms, the book uses a different synthetic dataset called wave. It has a single input feature and a continuous target variable, which is the response the algorithm tries to model. The code creates 40 points, and the plot puts the feature on the horizontal axis and the target on the vertical axis.

The target is plotted as a continuous quantity, not as a set of colored classes. Each dot still corresponds to one row: its horizontal position is its feature value, and its vertical position is its target value. Because the target is numeric, the plot shows how the response changes as the feature changes.

Both datasets come from the same mglearn module and both are generated by a make function, but their roles differ. The forge data asks which class a point belongs to. The wave data asks what numeric value goes with each feature value.

Simple datasets are chosen because they are easy to see. A printed page has two dimensions, so a dataset with two features can be drawn directly, with one feature per axis. Data with many more features is hard to show at all. Datasets with few features are called low-dimensional, and datasets with many features are called high-dimensional.