# Choosing Scale to Reveal Structure

## Contents

# 11.1. Choosing Scale to Reveal Structure¶

In Section 10.5, we first explored prices for houses sold in the San Francisico Bay Area between 2003 and 2009. Let’s revisit that example and take a look at the following histogram of prices.

```
px.histogram(sfh, x='price', width=350, height=250)
```

Is this plot accurate? Yes—it displays all the data. But most of the data are crammed into the left side of the plot, which makes it hard to use the plot to understand house prices.

Through data visualization, we want to reveal important features of the data
like the shape of a distribution and the relationship between two or more features.
As this example shows, after we produce an initial plot there are still other aspects we need to consider.
In this section, we cover *principles of scale* which help us decide how to
adjust the axis limits, place tick marks, and apply transformations.
We begin by examining when and how we might adjust a plot to reduce empty space; in other words we try to fill the data region of our plot with data.

## 11.1.1. Filling the Data Region¶

As we can see from the above plot of house prices, it’s hard to read a distribution when most of the data appear in a small portion of the plotting region. When this happens, we can’t clearly see important features about the data like whether there are multiple modes or skewness. A similar issue happens for scatter plots. When all the points are bunched together in a scatter plot, it’s hard to see nonlinearity.

This issue can happen when there are a few unusually large observations. In order to get a better view of the main portion of the data we can drop these observations from the plot by adjusting the x- or y-axis limits, or by removing outlier values from the data before plotting. In either case, we must mention this exclusion in the caption or on the plot itself.

Let’s use this idea to improve the plot of housing prices.

In the side-by-side plots below, we clip the data by changing the limits of the x-axis.
On the left, we’ve excluded houses that cost over \(\$2,000,000\).
The shape of the distribution for the bulk of the houses is much clearer in this plot.
For instance, we can more easily observe the skewness of the main mode and possible bimodality.
On the right, we separately show detail in the long right tail of the distribution.

```
right = px.histogram(sfh, x='price')
left = px.histogram(sfh, x='price')
fig = left_right(right, left, height=300)
fig.update_xaxes(range=[0, 2e6], row=1, col=1)
fig.update_xaxes(range=[2e6, 9e6], row=1, col=2)
fig.update_yaxes(range=[0, 10], row=1, col=2)
fig
#fig.update_xaxes(range=[0, 2e6], row=1, col=2)
```

We addressed the issue of scale in the original histogram of housing prices by making two plots, one for the bulk of the data and one for the tail. By choosing useful x- and y-axis limits for both plots, we can show additional useful information about the distribution.

Notice that the x-axis in the right plot includes 0, but the left plot begins its x-axis at \(\$2,000,000\). We consider when to include or exclude 0 on an axis next.

## 11.1.2. Including Zero¶

We often don’t need to include 0 on an axis, especially if including it makes it difficult to fill the data region. For instance, the scatter plot below shows the average longevity plotted against average height for dog breeds. (This dataset is provided by the American Kennel Club (AKC) and was first introduced in {numref}‘Chapter %s ch:eda’; it includes several features for 172 breeds.) The x-axis of the plot starts at 10cm since all dogs are at least that tall, and, similarly, the y-axis begins at 5 years.

```
fig = px.scatter(dogs, x='height', y='longevity',
title='Longevity vs. height for dog breeds',
width=350, height=250)
margin(fig, t=30)
```

There are a few cases where we usually want to include 0. For bar charts, including 0 is important so the heights of the bars directly relate to the data values. As an example, we’ve created two bar charts that compare the longevity of dog breeds below. The left plot includes 0, but the right plot doesn’t. It’s easy to incorrectly conclude from the right plot that medium-sized dogs live twice as long as large-sized dogs.

```
dogs_lon = dogs.groupby('size')['longevity'].mean().reset_index()
sml = {"size": ['small', 'medium', 'large']}
left = px.bar(dogs_lon, x='longevity', y='size', category_orders=sml)
right = px.bar(dogs_lon, x='longevity', y='size', category_orders=sml)
fig = left_right(left, right, height=250)
fig.update_xaxes(range=[7, 13], row=1, col=2)
fig.update_xaxes(title_text='Avg Longevity (yrs)')
fig.update_layout(yaxis_title="Size")
```

We also typically want to include zero when working with proportions, since proportions range from 0 to 1. The plot below shows the proportion of dogs of each dog group in our dataset.

```
size_props = ((dogs['group'].value_counts() / len(dogs))
.reset_index()
.rename(columns={'index': 'group', 'group': 'proportion'}))
size_props
fig = px.scatter(size_props, x='proportion', y='group', category_orders=sml,
width=350, height=250)
fig.update_traces(marker_size=15)
fig.update_xaxes(range=[0, 0.5])
```

In both of these plots, by including 0, you are making it easier for your reader to accurately compare the relative size of groups.

Earlier, when we adjusted axes, we essentially dropped data from our plotting region. While this is a useful strategy when a handful of observations are unusually large (or small), it is less effective with skewed distributions. In this situation, we often need to transform the data to gain a better view of its shape.

## 11.1.3. Revealing Shape Through Transformations¶

Another common way to adjust scale is to transform the data or the plot axes. We use transformations for skewed data so that it is easier to inspect the distribution. And, when the transformation produces a symmetric distribution, the symmetry carries with it useful properties in later modeling steps.

There are multiple ways to transform data, but the log-transform is especially useful. For instance, we’ve reproduced two histograms of SF house prices below. The left histogram is the original data. On the right, we’ve taken the log (base 10) of the prices before plotting.

```
sfl = sfh.assign(log_price=np.log10(sfh['price']))
orig = px.histogram(sfl, x='price', width=350, height=250)
logged = px.histogram(sfl, x='log_price',
nbins=100,
labels=dict(log_price='log10(price)'),
width=350, height=250)
fig = left_right(orig, logged)
fig.update_xaxes(title_text='price', row=1, col=1)
fig.update_xaxes(title_text='log10(price)', row=1, col=2)
```

The log transformation makes the distribution of prices more symmetric. Now, we can more easily see important features of the distribution, like the mode at around \(10^{5.85} ≈ \$700,000\) and the secondary mode near \(10^{5.55} ≈ \$350,000\).

The downside of using the log transform is that the actual values aren’t as intuitive—in this example, we needed to convert the values back to original dollars. Because of this, we often favor transforming the axis to use a log scale. When we do this, we don’t transform the data. Instead, we set the axis scale when we plot the code.

```
fig = px.histogram(sfh, x='price',
log_x=True,
histnorm='probability density',
width=350, height=250)
fig.update_traces(xbins_size=30_000)
```

The above histogram with log scaled x-axis shows a similar result to directly transforming the data. But since the axis is displayed using original units, we can directly read off its values.

The log transform can also reveal shape in scatter plots. Below, we’ve put the building size on the x-axis and the lot size on the y-axis. It’s hard to see the shape in this plot since many of the points are crammed into the bottom corner of the data region.

```
px.scatter(sfh, x='bsqft', y='lsqft',
width=350, height=250)
```

However, if we use a log scale for both x- and y-axes, the shape is much easier to see.

```
px.scatter(sfh, x='bsqft', y='lsqft',
log_x=True, log_y=True,
width=350, height=250)
```

With the transformed axes, we can see that the lot size increases roughly linearly with building size (on the log scale). The log transformation pulls large values–values that are orders of magnitude larger than others–in. This transformation can help fill the data region and uncover hidden structure as we saw for both the distribution of house price and the relationship between house size and lot size.

In addition to setting the limits of an axis and transforming an axis, we also want to consider the aspect ratio of the plot. That is, we want to adjust the length and width of the rectangular plot. That is, we can stretch or shrink a plot without changing its axes limits. This adjustment is called “banking”. In the next section, we explain how banking can help reveal relationships between features.

## 11.1.4. Banking to Decipher Relationships¶

With scatterplots, we try to choose a scale so that the relationship between the two features roughly follows a 45-degree line. This is called “banking to 45 degrees.” Banking to 45 degrees makes it easier for the reader to see shape and trends because our eyes can more easily pick up deviation this way. That is, it is much easier for us to see departures from a line when the data roughly fall along a 45-degree line within the plotting region. For instance, we’ve reproduced the plot that shows longevity of dog breeds against height. The plot has been banked to 45 degrees, and we can more easily see how the data roughly follow a line and where they deviate a bit at the extremes.

```
fig = px.scatter(dogs, x='height', y='longevity',
width=300, height=300)
#fig.update_yaxes(
# scaleanchor="x",
# scaleratio=10,
#)
fig
```

While banking to 45 degrees helps us see whether or not the data follow a linear relationship, when there is clear curvature we often find it hard to figure out what nonlinear relationship the data are showing us. When this happens, we try transforming features to linearize our scatter plot. We saw earlier that when we examining highly skewed data, the log transformation is often preferrable to simply changing the scale of the x-axis. The log transformation can also be useful in uncovering the general form of curvilinear relationships.

## 11.1.5. Revealing Relationships Through Straightening¶

We often use scatter plots to look at the relationship between two variables. For instance, in the plot below we’ve plotted heights against weights for the AKC dog breeds. We see that taller dogs weigh more, but this relationship isn’t linear.

```
px.scatter(dogs, x='height', y='weight',
width=350, height=250)
```

When it looks like two variables have a non-linear relationship, it’s useful to try applying a log scale the x-axis, y-axis, or both. We look for a linear relationship in the scatter plot with transformed axes. For instance, in the plot below, we applied a log scale to both x- and y-axes.

```
px.scatter(dogs, x='height', y='weight',
log_x=True, log_y=True,
width=350, height=350)
```

This plot shows a roughly linear relationship when both axes are on a log scale. In this case, we can say that there’s a polynomial relationship between dog height and weight. This makes sense intuitively—if a dog is taller, it’s probably proportionally both longer and wider so we might expect that the weight of a dog has a cubic relationship with its height.

In general, when we see a linear relationship after transforming one or both axes, we can use
Table 11.1 to reveal what relationship the original variables have.
We make these transformations because it is easier for us to see whether points fall along a line or not than to see if they follow a polynomial rather than an exponetial curve.

(The derivations of these relationships requires a bit of algebra, which
we leave to the exercises.)

x-axis |
y-axis |
Relationship |
---|---|---|

No transform |
No transform |
Linear: \( y = ax + b \) |

Log-scale |
No transform |
Log: \( y = a \log x + b \) |

No transform |
Log-scale |
Exponential: \( y = ba^x \) |

Log-scale |
Log-scale |
Polynomial: \( y = bx^a \) |

As Table 11.1 shows, the log transform can reveal several common types of relationships. Because of this, the log transform is considered the jackknife of transformations. As another, albeit artificial, example, the leftmost plot in Figure 11.1 reveals a curvilinear relationship between x and y. The middle plot show a different curvilinear relationship between log(y) and x; this plot also appears nonlinear. A further log transformation, at the far right in the figure, displays a plot of log(y) against log(x). This plot confirms that the data have a log-log (or polynomial) relationship because the transformed points fall along a line.

Adjusting scale is an important practice in data visualization, and in this section, we showed several approaches and when each approach is useful. When the data have a few unusually large values, we can clip the axis limits to “drop” these points and zoom in on the bulk of the data. Another approach to filling the data region when the data are highly skewed is to apply a log transformation. These transformations can help reveal features that are otherwise not apparent in the untransformed graphs. The log transformation is a very useful tool for uncovering the shape of a distribution and the form of a relationship. When we straighten a curve by using a log transformation on one or both axes, we can better assess the form of the original nonlinear relationship. And, changing the length and width of a scatter plot so that the data bank to 45 degrees helps us see curvature.

In the next section, we’ll look at principles of smoothing which we use when we need to visualize lots of data at once.