Plotting with Matplotlib

This page demonstrates the use of the Matplotlib plotting backend, the equivalent page demonstrating the Plotly backend may be found here

As we discovered in the Introduction, HoloViews allows plotting a variety of data types. Here we will use the sample data module and load the pandas and dask hvPlot API:

import numpy as np
import hvplot.pandas  # noqa
import hvplot.dask  # noqa

hvplot.extension('matplotlib')

As we learned the hvPlot API closely mirrors the Pandas plotting API, but instead of generating static images when used in a notebook, it uses HoloViews to generate either static or dynamically streaming Bokeh plots. Static plots can be used in any context, while streaming plots require a live Jupyter notebook, a deployed Bokeh Server app, or a deployed Panel app.

HoloViews provides an extensive, very rich set of objects along with a powerful set of operations to apply, as you can find out in the HoloViews User Guide. But here we will focus on the most essential mechanisms needed to make your data visualizable, without having to worry about the mechanics going on behind the scenes.

We will be focusing on two different datasets:

• A small CSV file of US crime data, broken down by state

• A larger Parquet-format file of airline flight data

The hvplot.sample_data module makes these datasets Intake data catalogue, which we can load either using pandas:

from hvplot.sample_data import us_crime, airline_flights

crime = us_crime.read()
print(type(crime))
crime.head()

<class 'pandas.core.frame.DataFrame'>

	Year	Population	Violent crime total	Murder and nonnegligent Manslaughter	Legacy rape /1	Revised rape /2	Robbery	Aggravated assault	Property crime total	Burglary	...	Violent Crime rate	Murder and nonnegligent manslaughter rate	Legacy rape rate /1	Revised rape rate /2	Robbery rate	Aggravated assault rate	Property crime rate	Burglary rate	Larceny-theft rate	Motor vehicle theft rate
0	1960	179323175	288460	9110	17190	NaN	107840	154320	3095700	912100	...	160.9	5.1	9.6	NaN	60.1	86.1	1726.3	508.6	1034.7	183.0
1	1961	182992000	289390	8740	17220	NaN	106670	156760	3198600	949600	...	158.1	4.8	9.4	NaN	58.3	85.7	1747.9	518.9	1045.4	183.6
2	1962	185771000	301510	8530	17550	NaN	110860	164570	3450700	994300	...	162.3	4.6	9.4	NaN	59.7	88.6	1857.5	535.2	1124.8	197.4
3	1963	188483000	316970	8640	17650	NaN	116470	174210	3792500	1086400	...	168.2	4.6	9.4	NaN	61.8	92.4	2012.1	576.4	1219.1	216.6
4	1964	191141000	364220	9360	21420	NaN	130390	203050	4200400	1213200	...	190.6	4.9	11.2	NaN	68.2	106.2	2197.5	634.7	1315.5	247.4

5 rows × 22 columns

Or using dask as a dask.DataFrame:

flights = airline_flights.to_dask().persist()
print(type(flights))
flights.head()

<class 'dask.dataframe.dask_expr._collection.DataFrame'>

	year	month	day	dayofweek	dep_time	crs_dep_time	arr_time	crs_arr_time	carrier	flight_num	...	taxi_in	taxi_out	cancelled	cancellation_code	diverted	carrier_delay	weather_delay	nas_delay	security_delay	late_aircraft_delay
0	2008	11	15	6	1411.0	1420	1535.0	1546	OO	4391	...	5.0	11.0	0	None	0	NaN	NaN	NaN	NaN	NaN
1	2008	11	28	5	1222.0	1230	1345.0	1356	OO	4391	...	5.0	15.0	0	None	0	NaN	NaN	NaN	NaN	NaN
2	2008	11	22	6	1414.0	1420	1540.0	1546	OO	4391	...	5.0	10.0	0	None	0	NaN	NaN	NaN	NaN	NaN
3	2008	11	15	6	1304.0	1305	1507.0	1519	OO	4392	...	10.0	9.0	0	None	0	NaN	NaN	NaN	NaN	NaN
4	2008	11	22	6	1323.0	1305	1536.0	1519	OO	4392	...	5.0	21.0	0	None	0	0.0	0.0	0.0	0.0	17.0

5 rows × 29 columns

The plot interface

The dask.dataframe.DataFrame.hvplot, pandas.DataFrame.hvplot and intake.DataSource.plot interfaces (and Series equivalents) from HvPlot provide a powerful high-level API to generate complex plots. The .hvplot API can be called directly or used as a namespace to generate specific plot types.

The plot method

The most explicit way to use the plotting API is to specify the names of columns to plot on the x- and y-axis respectively:

crime.hvplot.line(x='Year', y='Violent Crime rate')

As you’ll see in more detail below, you can choose which kind of plot you want to use for the data:

crime.hvplot(x='Year', y='Violent Crime rate', kind='scatter')

To group the data by one or more additional columns, specify an additional by variable. As an example here we will plot the departure delay (‘depdelay’) as a function of ‘distance’, grouping the data by the ‘carrier’. There are many available carriers, so we will select only two of them so that the plot is readable:

flight_subset = flights[flights.carrier.isin(['OH', 'F9'])]
flight_subset.hvplot(x='distance', y='depdelay', by='carrier', kind='scatter', alpha=0.2, persist=True)

Here we have specified the x axis explicitly, which can be omitted if the Pandas index column is already the desired x axis. Similarly, here we specified the y axis; by default all of the non-index columns would be plotted (which would be a lot of data in this case). If you don’t specify the ‘y’ axis, it will have a default label named ‘value’, but you can then provide a y axis label explicitly using the value_label option.

Putting all of this together we will plot violent crime, robbery, and burglary rates on the y-axis, specifying ‘Year’ as the x, and relabel the y-axis to display the ‘Rate’.

crime.hvplot(x='Year', y=['Violent Crime rate', 'Robbery rate', 'Burglary rate'],
             value_label='Rate (per 100k people)')

The hvplot namespace

Instead of using the kind argument to the plot call, we can use the hvplot namespace, which lets us easily discover the range of plot types that are supported. Use tab completion to explore the available plot types:

crime.hvplot.<TAB>

Plot types available include:

• .area(): Plots a area chart similar to a line chart except for filling the area under the curve and optionally stacking

• .bar(): Plots a bar chart that can be stacked or grouped

• .bivariate(): Plots 2D density of a set of points

• .box(): Plots a box-whisker chart comparing the distribution of one or more variables

• .heatmap(): Plots a heatmap to visualizing a variable across two independent dimensions

• .hexbins(): Plots hex bins

• .hist(): Plots the distribution of one or histograms as a set of bins

• .kde(): Plots the kernel density estimate of one or more variables.

• .line(): Plots a line chart (such as for a time series)

• .scatter(): Plots a scatter chart comparing two variables

• .step(): Plots a step chart akin to a line plot

• .table(): Generates a SlickGrid DataTable

• .violin(): Plots a violin plot comparing the distribution of one or more variables using the kernel density estimate

Area

Like most other plot types the area chart supports the three ways of defining a plot outlined above. An area chart is most useful when plotting multiple variables in a stacked chart. This can be achieve by specifying x, y, and by columns or using the columns and index/use_index (equivalent to x) options:

crime.hvplot.area(x='Year', y=['Robbery', 'Aggravated assault'])

We can also explicitly set stacked to False and define an alpha value to compare the values directly:

crime.hvplot.area(x='Year', y=['Aggravated assault', 'Robbery'], stacked=False, alpha=0.4)

Another use for an area plot is to visualize the spread of a value. For instance using the flights dataset we may want to see the spread in mean delay values across carriers. For that purpose we compute the mean delay by day and carrier and then the min/max mean delay for across all carriers. Since the output of hvplot is just a regular holoviews object, we can use the overlay operator (*) to place the plots on top of each other.

delay_min_max = flights.groupby(['day', 'carrier'])['carrier_delay'].mean().groupby('day').agg(['min', 'max'])
delay_mean = flights.groupby('day')['carrier_delay'].mean()

delay_min_max.hvplot.area(x='day', y='min', y2='max', alpha=0.2) * delay_mean.hvplot()

Bars

In the simplest case we can use .hvplot.bar to plot x against y. We’ll use rot=90 to rotate the tick labels on the x-axis making the years easier to read:

crime.hvplot.bar(x='Year', y='Violent Crime rate', rot=90, width=900)

If we want to compare multiple columns instead we can set y to a list of columns. Using the stacked option we can then compare the column values more easily:

crime.hvplot.bar(x='Year', y=['Violent crime total', 'Property crime total'],
                 stacked=True, rot=90, width=900, legend='top_left')

Scatter

The scatter plot supports many of the same features as the other chart types we have seen so far but can also be colored by another variable using the c option.

crime.hvplot.scatter(x='Violent Crime rate', y='Burglary rate', c='Year')

Anytime that color is being used to represent a dimension, the cmap option can be used to control the colormap that is used to represent that dimension. Additionally, the colorbar can be disabled using colorbar=False.

Step

A step chart is very similar to a line chart but instead of linearly interpolating between samples the step chart visualizes discrete steps. The point at which to step can be controlled via the where keyword allowing 'pre', 'mid' (default) and 'post' values:

crime.hvplot.step(x='Year', y=['Robbery', 'Aggravated assault'])

HexBins

You can create hexagonal bin plots with the hexbin method. Hexbin plots can be a useful alternative to scatter plots if your data are too dense to plot each point individually. Since these data are not regularly distributed, we’ll use the logz option to map z-axis (color) to a log scale colorbar.

# flights.hvplot.hexbin(x='airtime', y='arrdelay', width=600, height=500, logz=True);

Output suppressed as this is not currently supported with the Matplotlib backend and doesn't display any plot.

Bivariate

You can create a 2D density plot with the bivariate method. Bivariate plots can be a useful alternative to scatter plots if your data are too dense to plot each point individually.

crime.hvplot.bivariate(x='Violent Crime rate', y='Burglary rate', width=600, height=500)

HeatMap

A HeatMap lets us view the relationship between three variables, so we specify the ‘x’ and ‘y’ variables and an additional ‘C’ variable. Additionally we can define a reduce_function that computes the values for each bin from the samples that fall into it. Here we plot the ‘depdelay’ (i.e. departure delay) for each day of the month and carrier in the dataset:

flights.compute().hvplot.heatmap(x='day', y='carrier', C='depdelay', reduce_function=np.mean).opts(show_values=False)

Tables

Unlike all other plot types, a table only supports one signature: either all columns are plotted, or a subset of columns can be selected by defining the columns explicitly:

crime.hvplot.table(columns=['Year', 'Population', 'Violent Crime rate'], width=400)

Distributions

Plotting distributions differs slightly from other plots since they plot only one variable in the simple case rather than plotting two or more variables against each other. Therefore when plotting these plot types no index or x value needs to be supplied. Instead:

1. Declare a single y variable, e.g. source.plot.hist(variable), or

2. Declare a y variable and by variable, e.g. source.plot.hist(variable, by='Group'), or

3. Declare columns or plot all columns, e.g. source.plot.hist() or source.plot.hist(columns=['A', 'B', 'C'])

Histogram

The Histogram is the simplest example of a distribution; often we simply plot the distribution of a single variable, in this case the ‘Violent Crime rate’. Additionally we can define a range over which to compute the histogram and the number of bins using the bin_range and bins arguments respectively:

crime.hvplot.hist(y='Violent Crime rate')

Or we can plot the distribution of multiple columns:

columns = ['Violent Crime rate', 'Property crime rate', 'Burglary rate']
crime.hvplot.hist(y=columns, bins=50, alpha=0.5, legend='top', height=400)

We can also group the data by another variable. Here we’ll use subplots to split each carrier out into its own plot:

flight_subset = flights[flights.carrier.isin(['AA', 'US', 'OH'])]
# flight_subset.hvplot.hist('depdelay', by='carrier', bins=20, bin_range=(-20, 100), width=300, subplots=True);

Output suppressed as this is not currently supported with the Matplotlib backend and raises an error. Tip: execute flight_subset.compute() before plotting to avoid the error.

KDE (density)

You can also create density plots using hvplot.kde() or hvplot.density():

crime.hvplot.kde(y='Violent Crime rate')

Comparing the distribution of multiple columns is also possible:

columns=['Violent Crime rate', 'Property crime rate', 'Burglary rate']
crime.hvplot.kde(y=columns, alpha=0.5, value_label='Rate', legend='top_right')

The hvplot.kde also supports the by keyword:

flight_subset = flights[flights.carrier.isin(['AA', 'US', 'OH'])]
flight_subset.hvplot.kde('depdelay', by='carrier', xlim=(-20, 70), width=300, subplots=True)

Box-Whisker Plots

Just like the other distribution-based plot types, the box-whisker plot supports plotting a single column:

crime.hvplot.box(y='Violent Crime rate')

It also supports multiple columns and the same options as seen previously (legend, invert, value_label):

columns=['Burglary rate', 'Larceny-theft rate', 'Motor vehicle theft rate',
         'Property crime rate', 'Violent Crime rate']
crime.hvplot.box(y=columns, group_label='Crime', legend=False, value_label='Rate (per 100k)', invert=True)

Lastly, it also supports using the by keyword to split the data into multiple subsets:

flight_subset = flights[flights.carrier.isin(['AA', 'US', 'OH'])]
# flight_subset.hvplot.box('depdelay', by='carrier', ylim=(-10, 70));

Output suppressed as this is not currently supported with the Matplotlib backend and displays an empty plot.

Composing Plots

One of the core strengths of HoloViews is the ease of composing different plots. Individual plots can be composed using the * and + operators, which overlay and compose plots into layouts respectively. For more information on composing objects, see the HoloViews User Guide.

By using these operators we can combine multiple plots into composite plots. A simple example is overlaying two plot types:

crime.hvplot(x='Year', y='Violent Crime rate') * crime.hvplot.scatter(x='Year', y='Violent Crime rate', c='k')

We can also lay out different plots and tables together:

(crime.hvplot.bar(x='Year', y='Violent Crime rate', rot=90, width=550) +
 crime.hvplot.table(['Year', 'Population', 'Violent Crime rate'], width=420))

We can further enhance the plot by using a steps chart instead of a bar chart to avoid visual clutter and better represent the continuous nature of time series data. Additionally, we can display a table summarizing decade-averaged statistics to provide a quick overview of the data. The sublabel format and size can be customized, and if not needed, can be removed by setting sublabel_format="" in opts.

tabledata = (
    crime.assign(
        **{
            "Decade": (crime.Year // 10) * 10,  # Calculate decade for each year
            "Population [mi]": crime.Population / 1e6,  # Population in millions
        }
    )
    .groupby("Decade")  # Group data by decade
    .mean()  # Calculate mean values for each decade group
    [["Population [mi]", "Violent Crime rate"]]  # Select desired columns
    .astype("int")  # Convert selected columns to integers
)

(
    crime.hvplot.step(x='Year', y='Violent Crime rate', rot=90, width=550)+
    tabledata.hvplot.table(['Decade', 'Population [mi]', 'Violent Crime rate'], width=15, height=10)
).opts(sublabel_format="{alpha})", sublabel_size=14)

Large data

The previous examples summarized the fairly large airline dataset using statistical plot types that aggregate the data into a feasible subset for plotting. We can instead aggregate the data directly into the viewable image using datashader, which provides a rendering of the entire set of raw data available (as far as the resolution of the screen allows). Here we plot the ‘airtime’ against the ‘distance’:

flights.hvplot.scatter(x='distance', y='airtime', datashade=True)

Groupby

Thanks to the ability of HoloViews to explore a parameter space with a set of widgets we can apply a groupby along a particular column or dimension. For example we can view the distribution of departure delays by carrier grouped by day, allowing the user to choose which day to display:

flights.hvplot.violin(y='depdelay', by='carrier', groupby='dayofweek', ylim=(-20, 60), height=500)

This user guide merely provided an overview over the available plot types; to see a detailed description on how to customize plots see the Plotting Options reference.

This web page was generated from a Jupyter notebook and not all interactivity will work on this website.