Using scikit-learn for Machine Learning

Imagine you want to do machine learning in Python on a computer that has no scientific Python tools yet. You would need a numerical computing library, a library for scientific algorithms, plotting tools, data-handling tools, an interactive environment for experiments, and finally scikit-learn itself. Each of these could be installed on its own, but the choices and the order of installation quickly become a chore.

This article looks at one way around that problem: a single Python distribution called Anaconda, which arrives with most of these tools already in place. As you read, keep one question in mind: what does it mean for a distribution to already contain the packages you need, and what does that save you from doing?

The word distribution is the key idea. When you install Anaconda, you are not installing one program. You are installing a container that holds several components at once, so that every component beneath it arrives as part of that single installation. Because it is offered on Mac OS, Windows, and Linux, the same bundle idea applies across the three major desktop platforms.

Think of the installation as a set of layers. The distribution sits at the top, and the packages that scikit-learn depends on and works alongside sit beneath it. The source names seven packages that come with Anaconda: NumPy, SciPy, matplotlib, pandas, IPython, Jupyter Notebook, and scikit-learn. These cover different roles, from numerical arrays to plotting to interactive work, and all of them are present after one installation.

Compare the two routes to a working machine learning setup. In the manual route, you must first decide on each library, work out an order in which to install them, and then install scikit-learn last. In the Anaconda route, you install one distribution, and the packages it contains are already in place. The steps differ mainly in how many separate decisions you have to make.

Anaconda also includes the Intel MKL library. The source describes it as commercial software that is provided for free within Anaconda. You do not choose to add it, and using it happens automatically when Anaconda is installed.

The reason MKL matters to a scikit-learn user is performance. The source states that using MKL can give significant speed improvements for many algorithms in scikit-learn. Read the chain from left to right: an algorithm you call in scikit-learn runs, and the Intel MKL library, included in Anaconda, works on the algorithms that scikit-learn runs. The source does not give further detail on the internal mechanism, so this article does not go beyond that relationship.

The source gives one clear rule: Anaconda is suggested for people who do not already have an installation of the scientific Python packages. You can turn that into a short sequence of checks. Work through them in order and stop at the first answer that settles the matter.