Multi-View Composition

We will be visualizing weather statistics for the U.S. cities of Seattle and New York. Let's load the dataset and peek at the first and last 10 rows:

We will create multi-view displays to examine weather within and across the cities.

One of the most common ways of combining multiple charts is to layer marks on top of each other. If the underlying scale domains are compatible, we can merge them to form shared axes. If either of the x or y encodings is not compatible, we might instead create a dual-axis chart, which overlays marks using separate scales and axes.

Faceting involves subdividing a dataset into groups and creating a separate plot for each group. In earlier notebooks, we learned how to create faceted charts using the row and column encoding channels. We'll first review those channels and then show how they are instances of the more general facet operator.

Let's start with a basic histogram of maximum temperature values in Seattle:

How does this temperature profile change based on the weather of a given day – that is, whether there was drizzle, fog, rain, snow, or sun?

Let's use the column encoding channel to facet the data by weather type. We can also use color as a redundant encoding, using a customized color range:

Unsurprisingly, those rare snow days center on the coldest temperatures, followed by rainy and foggy days. Sunny days are warmer and, despite Seattle stereotypes, are the most plentiful. Though as any Seattleite can tell you, the drizzle occasionally comes, no matter the temperature!

In addition to row and column encoding channels within a chart definition, we can take a basic chart definition and apply faceting using an explicit facet operator.

Let's recreate the chart above, but this time using facet. We start with the same basic histogram definition, but remove the data source, filter transform, and column channel. We can then invoke the facet method, passing in the data and specifying that we should facet into columns according to the weather field. The facet method accepts both row and column arguments. The two can be used together to create a 2D grid of faceted plots.

Finally we include our filter transform, applying it to the top-level faceted chart. While we could apply the filter transform to the histogram definition as before, that is slightly less efficient. Rather than filter out "New York" values within each facet cell, applying the filter to the faceted chart lets Vega-Lite know that we can filter out those values up front, prior to the facet subdivision.

Given all the extra code above, why would we want to use an explicit facet operator? For basic charts, we should certainly use the column or row encoding channels if we can. However, using the facet operator explicitly is useful if we want to facet composed views, such as layered charts.

Let's revisit our layered temperature plots from earlier. Instead of plotting data for New York and Seattle in the same plot, let's break them up into separate facets. The individual chart definitions are nearly the same as before: one area chart and one line chart. The only difference is that this time we won't pass the data directly to the chart constructors; we'll wait and pass it to the facet operator later. We can layer the charts much as before, then invoke facet on the layered chart object, passing in the data and specifying column facets based on the location field:

The faceted charts we have seen so far use the same axis scale domains across the facet cells. This default of using shared scales and axes helps aid accurate comparison of values. However, in some cases you may wish to scale each chart independently, for example if the range of values in the cells differs significantly.

Similar to layered charts, faceted charts also support resolving to independent scales or axes across plots. Let's see what happens if we call the resolve_axis method to request independent y-axes:

The chart above looks largely unchanged, but the plot for Seattle now includes its own axis.

What if we instead call resolve_scale to resolve the underlying scale domains?

Now we see facet cells with different axis scale domains. In this case, using independent scales seems like a bad idea! The domains aren't very different, and one might be fooled into thinking that New York and Seattle have similar maximum summer temperatures.

To borrow a cliché: just because you can do something, doesn't mean you should...

Faceting creates small multiple plots that show separate subdivisions of the data. However, we might wish to create a multi-view display with different views of the same dataset (not subsets) or views involving different datasets.

Altair provides concatenation operators to combine arbitrary charts into a composed chart. The hconcat operator (shorthand | ) performs horizontal concatenation, while the vconcat operator (shorthand &) performs vertical concatenation.

Let's start with a basic line chart showing the average maximum temperature per month for both New York and Seattle, much like we've seen before:

What if we want to compare not just temperature over time, but also precipitation and wind levels?

Let's create a concatenated chart consisting of three plots. We'll start by defining a "base" chart definition that contains all the aspects that should be shared by our three plots. We can then modify this base chart to create customized variants, with different y-axis encodings for the temp_max, precipitation, and wind fields. We can then concatenate them using the pipe (|) shorthand operator:

Alternatively, we could use the more explicit alt.hconcat() method in lieu of the pipe | operator. Try rewriting the code above to use hconcat instead.

Vertical concatenation works similarly to horizontal concatenation. Using the & operator (or alt.vconcat method), modify the code to use a vertical ordering instead of a horizontal ordering.

Finally, note that horizontal and vertical concatenation can be combined. What happens if you write something like (temp | precip) & wind?

Aside: Note the importance of those parentheses... what happens if you remove them? Keep in mind that these overloaded operators are still subject to Python's operator precendence rules, and so vertical concatenation with & will take precedence over horizontal concatenation with |!

As we will revisit later, concatenation operators let you combine any and all charts into a multi-view dashboard!

The concatenation operators above are quite general, allowing arbitrary charts to be composed. Nevertheless, the example above was still a bit verbose: we have three very similar charts, yet have to define them separately and then concatenate them.

For cases where only one or two variables are changing, the repeat operator provides a convenient shortcut for creating multiple charts. Given a template specification with some free variables, the repeat operator will then create a chart for each specified assignment to those variables.

Let's recreate our concatenation example above using the repeat operator. The only aspect that changes across charts is the choice of data field for the y encoding channel. To create a template specification, we can use the repeater variable alt.repeat('column') as our y-axis field. This code simply states that we want to use the variable assigned to the column repeater, which organizes repeated charts in a horizontal direction. (As the repeater provides the field name only, we have to specify the field data type separately as type='quantitative'.)

We then invoke the repeat method, passing in data field names for each column:

Repetition is supported for both columns and rows. What happens if you modify the code above to use row instead of column?

We can also use row and column repetition together! One common visualization for exploratory data analysis is the scatter plot matrix (or SPLOM). Given a collection of variables to inspect, a SPLOM provides a grid of all pairwise plots of those variables, allowing us to assess potential associations.

Let's use the repeat operator to create a SPLOM for the temp_max, precipitation, and wind fields. We first create our template specification, with repeater variables for both the x- and y-axis data fields. We then invoke repeat, passing in arrays of field names to use for both row and column. Altair will then generate the cross product (or, Cartesian product) to create the full space of repeated charts:

Looking at these plots, there does not appear to be a strong association between precipitation and wind, though we do see that extreme wind and precipitation events occur in similar temperature ranges (~5-15° C). However, this observation is not particularly surprising: if we revisit our histogram at the beginning of the facet section, we can plainly see that the days with maximum temperatures in the range of 5-15° C are the most commonly occurring.

Modify the code above to get a better understanding of chart repetition. Try adding another variable (temp_min) to the SPLOM. What happens if you rearrange the order of the field names in either the row or column parameters for the repeat operator?

Finally, to really appreciate what the repeat operator provides, take a moment to imagine how you might recreate the SPLOM above using only hconcat and vconcat!

Together, the composition operators layer, facet, concat, and repeat form a view composition algebra: the various operators can be combined to construct a variety of multi-view visualizations.

As an example, let's start with two basic charts: a histogram and a simple line (a single rule mark) showing a global average.

We can then combine the two charts using a layer operator, and then repeat that layered chart to show histograms with overlaid averages for multiple fields:

Focusing only on the multi-view composition operators, the model for the visualization above is:

repeat(column=[...]) |- layer |- basic1 |- basic2

Now let's explore how we can apply all the operators within a final dashboard that provides an overview of Seattle weather. We'll combine the SPLOM and faceted histogram displays from earlier sections with the repeated histograms above:

The full composition model for this dashboard is:

vconcat |- hconcat | |- repeat(row=[...], column=[...]) | | |- splom base chart | |- repeat(row=[...]) | |- layer | |- dateHist base chart 1 | |- dateHist base chart 2 |- facet(column='weather') |- tempHist base chart

Phew! The dashboard also includes a few customizations to improve the layout:

Try removing or modifying any of these adjustments and see how the dashboard layout responds!

This dashboard can be reused to show data for other locations or from other datasets. Update the dashboard to show weather patterns for New York instead of Seattle.

For more details on multi-view composition, including control over sub-plot spacing and header labels, see the Altair Compound Charts documentation.

Now that we've seen how to compose multiple views, we're ready to put them into action. In addition to statically presenting data, multiple views can enable interactive multi-dimensional exploration. For example, using linked selections we can highlight points in one view to see corresponding values highlight in other views.

In the next notebook, we'll examine how to author interactive selections for both individual plots and multi-view compositions.