Chapter 1: Examining the Distribution of Features and Targets

Machine learning writing and instruction are often algorithm-focused. Sometimes, this gives the impression that all we have to do is choose the right model and that organization-changing insights will follow. But the best place to begin a machine learning project is with an understanding of how the features and targets we will use are distributed.

It is important to make room for the same kind of learning from data that has been central to our work as analysts for decades – studying the distribution of variables, identifying anomalies, and examining bivariate relationships – even as we focus more and more on the accuracy of our predictions.

We will explore tools for doing so in the first three chapters of this book, while also considering implications for model building.

In this chapter, we will use common NumPy and pandas techniques to get a better sense of the attributes of our data. We want to know how key features are distributed before we do any predictive analyses. We also want to know the central tendency, shape, and spread of the distribution of each continuous feature and have a count for each value for categorical features. We will take advantage of very handy NumPy and pandas tools for generating summary statistics, such as the mean, min, and max, as well as standard deviation.

After that, we will create visualizations of key features, including histograms and boxplots, to give us a better sense of the distribution of each feature than we can get by just looking at summary statistics. We will hint at the implications of feature distribution for data transformation, encoding and scaling, and the modeling that we will be doing in subsequent chapters with the same data.

Specifically, in this chapter, we are going to cover the following topics:

• Subsetting data
• Generating frequencies for categorical features
• Generating summary statistics for continuous features
• Identifying extreme values and outliers in univariate analysis
• Using histograms, boxplots, and violin plots to examine the distribution of continuous features