Visualizing Feature Distributions with Histograms

Count data appears whenever you record how many times something happened: visits to a page, items bought in a store, or appearances of a word. Features built this way share a recognizable pattern. Most values are small, a few values are very large, and the bulk of the data sits on one side of the range. When you draw a histogram of such a feature, the tall bars cluster near the low end and a long, thin tail stretches out to the right.

This shape matters for modeling. A linear model fitted directly to a feature like this can struggle, because a handful of extreme values pulls the fitted line toward themselves, and the model then captures the main relationship poorly. The source example that this article follows makes the point concretely: a Ridge regression model reached only a relatively low R squared score on its data, meaning it did not capture the relationship between the feature and the target well.

One way to address a skewed count feature is to transform it with a logarithm before fitting the model. This article shows exactly how that transformation is written, what it changes in the data, and how to check the change with a histogram. The central skill is tracing what each step does to the values, so the sections below move through the operations one at a time.

The transformation is a single line of NumPy applied to the training and test feature matrices. Before you read the explanation, try to predict what the expression does to a zero. The question is posed directly to you below, so write down your answer first.

The source gives the reason for the shift directly. The value 0 appears in the data, and the logarithm is not defined at 0. Because of this, applying the logarithm to the raw values is not an option. The transformation is therefore written as $\log(X + 1)$, which first adds 1 to each value and then takes the logarithm.

The shift guarantees that the argument of the logarithm is never zero for a count of zero, since 0 + 1 is 1. The original count of 0 is therefore mapped to a valid input rather than an undefined one. The shift also keeps every non-negative count within the domain of the logarithm, which is why the same expression works across the whole column without special cases.

A histogram shows how many values fall into each range, so it is a natural way to confirm that the transformation did what it was meant to do. The source plots the feature after the transformation, with a label on each axis and 25 bins. Plotting the transformed column in the same way lets you compare its shape with a histogram of the raw column.

The source describes the effect in two parts. First, the distribution becomes less asymmetrical: the lopsided pile of small values and the long tail of large values become more balanced. Second, the very large outliers that stretched the range are reduced. A logarithm compresses large values much more than small ones, so the long tail of big counts ends up closer to the bulk of the data. In the example above, the raw count of 99 sits nearly a hundred units away from the count of 1, while the transformed values for 1 and 99 are separated by about 3.9 units on the log scale.

A transformation usually sits inside a longer pipeline: the data is loaded, the transformation is applied, and the result is plotted or fed to a model. When a histogram looks wrong, the useful question is which stage first produced the unexpected result. Tracing the stages in order lets you find that point, rather than guessing from the final plot.