Preparing Data for scikit-learn

Before any machine learning model can learn from data, that data has to be in a form the library can accept. In scikit-learn, that form is the NumPy array. NumPy is one of the fundamental packages for scientific computing in Python, and it provides functionality for multidimensional arrays, high-level mathematical functions, and pseudorandom number generators. This article focuses on one piece of that picture: how you create a NumPy array, what kind of object you have once you do, and how to reason about it when something goes wrong.

The approach here is code first. You will read a short program, predict what it prints, and only then use the explanation to check your prediction. Pay attention to the gap between your prediction and the real output, because that gap is where understanding forms.

Consider this program. It passes a list of two lists to np.array and then prints the result and its shape. Read it before you look at the output.

The argument to np.array in that program is a list whose two items are themselves lists. Each inner list becomes one row of the resulting two-dimensional array. The outer structure gives the array its two rows, and each row contributes three elements. So the value 1 sits in the first row and first column, while 6 sits in the second row and third column.

Indexing in Python starts at zero, so positions are labeled with zero-based row and column numbers. The diagram below shows each value in its cell, with row 0 being the first row and column 0 being the first column.

The shape is the object's way of reporting its size. For the array above, the shape is (2, 3). The first number counts the rows, which is the length of the outer list, and the second number counts the columns, which is the length of each inner list. Axis 0 runs down the rows and axis 1 runs across the columns, so the axis numbers line up with the shape tuple position by position.

The word n-dimensional means the same class can describe a one-dimensional list of values, a two-dimensional table like the one above, or higher-dimensional containers. The two-dimensional case is the one you will meet most often when preparing rows of samples and their features.

A central rule of the ndarray class is that all elements of a NumPy array must be of the same type. A plain Python list has no such requirement: one list can hold an integer, a string, and a decimal number side by side. The array is a more constrained container, and that constraint is what lets it act as the common data structure for scikit-learn.

scikit-learn takes in data in the form of NumPy arrays. This is a requirement, not a stylistic preference: any data you want to use with scikit-learn has to be converted to a NumPy array first. Because of this, the array is the fundamental data structure of the library. When you load or build data for a model, the conversion step belongs in your workflow before you call any scikit-learn function.

When a result differs from your prediction, the location of the difference usually points to one of three places: the input list you handed to np.array, the call that created the array, or the step that printed or used it. Trace the program in that order, comparing what you expected at each line with what the code actually produces. The earliest line where expectation and output part ways is the place to investigate.