If the repository does not contain the plot type you need, please consider submitting a pull request to add it.
- Prefer horizontal layout over vertical layouts. Differences in width are easier to perceive than differences in height.
- Show the variability in your results. Results are rarely perfectly consistent.
- Label your axis.
- include the unit in the axis label (e.g.,
[s]for seconds). - Avoid large numbers on axis labels (e.g., use
1 MB, not1000 KB). - Make sure your axes starts at 0 to fairly represent the differences in values.
- Use the right amount of grid lines. Choose to use major and/or minor grid lines depending on your plot. Too many grid lines clutters the plot. Too few grid lines makes the plot hard to read.
The code snippets shown on this page require one or more of the libraries imported below.
library(tidyverse) # import ggplot, dplyr, readr, etc.
library(data.table) # data analysis
library(zoo) # for working with time series data
library(coronavirus) # dataset for example plots
theme_set(theme_bw()) # set black and white theme for plots.
library(gghighlight) # for highlighting in plots
library(cowplot) # convenience functions for plotting
library(knitr) # literate programming library
# saves plot and removes surrounding whitespace
saveplot <- function(filename, ...) {
ggsave2(filename, ...)
knitr::plot_crop(filename)
} # load data set
iris %>%
# group by species for summary on next line
group_by(Species) %>%
# calculate the arithmetic mean and the 5th and 95th percentile
summarize(
length = mean(Sepal.Length),
p5 = quantile(Sepal.Length,0.05),
p95 = quantile(Sepal.Length, 0.95)) %>%
# plot data. 'fill=Species' adds the colors
ggplot(aes(y=length, x=Species, fill=Species)) +
# create bar chart where data contains value to plot (mean, 5th percentile, 95th percentile)
geom_col() +
# add whiskers with 5th percentile and 95th percentile
geom_errorbar(aes(ymin=p5, ymax=p95), width=.4) +
# make sure plot starts at 0 cm
ylim(0,NA) +
# add pretty axis labels
labs(x= "", y="sepal length [cm]") +
# modify plot theme
theme_half_open() +
# add major grid lines on the horizontal axis
background_grid(major = "x") +
# remove legend
theme(legend.position = "none") +
# make bars horizontal, not vertical
coord_flip()
Bar charts are useful for comparing scalar values such as the makespan of a workload when using different schedulers, or the latency of request for different network configurations. However, when there is variability in your results (i.e., rerunning your experiments results in different values), consider using a box plot.
We plot the sepal length for all flowers in our data. The vertical axis shows the different flower types, and the horizontal axis shows the average sepal length, as well as the 5th and 95th percentiles.
# load data set
iris %>%
# plot data. 'fill=Species' adds the different colors
ggplot(aes(y=Sepal.Length, x=Species, fill=Species)) +
# create box plot
geom_boxplot() +
# calculate the arithmetic mean, for each species. Show in plot using a white circle with black border
stat_summary(fun=mean, geom="point", shape=21, size=2, color="black", fill="white") +
# make sure the horizontal axis starts at 0
ylim(0,NA) +
# add pretty labels
labs(x= "", y="sepal length [cm]") +
# set plot theme
theme_half_open() +
# add major grid lines for horizontal axis
background_grid(major = "x") +
# remove legend
theme(legend.position = "none") +
# make boxes horizontal, not vertical
coord_flip()
Whenever your scalar results show variability, it is good practice to consider using a box plot. The box plot is generally better at showing variability than a bar chart with whiskers because a box plot shows more (well known) statistical properties. The black line inside each box shows the median value, while the left and right side of the box show the 25th and 75th percentile respectively. Just as with the bar chart, the meaning of the whiskers in a box plot can vary. Typically, the whiskers extend to values that are within 1.5 times the inter-quartile range (IQR), i.e., within 1.5 times the width of the box. Values outside this range are outliers and are typically depicted individually as dots (in black). Alternative uses of the whiskers include, but are not limited to, the 5th and 95th percentile, and the minimum and maximum value.
Although the box plot shows several useful statistical properties, it does not show the commonly reported arithmetic mean (i.e., average) of a distribution. Unfortunately, there is no standardized way of showing the mean in a box plot. We find that a simple marker, distinctly different from the type of marker used for outliers, typically works well. In the example above we use a white dot with a black border to indicate the mean.
Line charts are useful for depicting events over time. The examples below show the number of COVID-19 cases reported over time in several places across the globe.
coronavirus %>%
# filter data
filter(type == "confirmed") %>%
filter(cases > 0) %>%
# create new column
mutate(place = ifelse(country == "US", "US", "other")) %>%
# order items in legend: first "US", then "other"
mutate(place = factor(place, levels = c("US", "other"))) %>%
# group and sum cases in US and elsewhere
group_by(date, place) %>%
summarise(cases = sum(cases), .groups='drop') %>%
ggplot(aes(x=date, y=cases, color=place)) +
geom_line() +
theme_half_open() +
scale_y_continuous(labels = function(x) floor(x/1000)) +
ylab(bquote("\u00D7"*10^3~"number of cases")) + background_grid(major = "y") +
theme(legend.position = "bottom")
A very simple line chart showing events over time.
coronavirus %>%
filter(type == "confirmed") %>%
filter(cases > 0) %>%
mutate(place = ifelse(country == "US", "US", "other")) %>%
# order items in legend: first "US", then "other"
mutate(place = factor(place, levels = c("US", "other"))) %>%
group_by(date, place) %>%
summarise(cases = sum(cases), .groups='drop') %>%
group_by(place) %>%
# create column with cumulative value
mutate(ccases = cumsum(cases)) %>%
ggplot(aes(x=date, y=ccases, color=place)) +
geom_line()+
theme_half_open() +
scale_y_continuous(labels = function(x) floor(x/1000000)) +
ylab(bquote("\u00D7"*10^6~"cumulative number of cases")) + background_grid(major = "y") +
theme(legend.position = "bottom")
A line chart can become difficult to read when the curves are highly jittery. In that case, it may help to plot cumulative values. However, this will make it more difficult to observe interesting events (e.g., spikes) in the data.
coronavirus %>%
filter(type == "confirmed") %>%
filter(cases > 0) %>%
mutate(place = ifelse(country == "US", "US", "other")) %>%
# order items in legend: first "US", then "other"
mutate(place = factor(place, levels = c("US", "other"))) %>%
group_by(date, place) %>%
summarise(cases = sum(cases), .groups="drop") %>%
group_by(place) %>%
# add column with sliding window average, window size = 7
mutate(meancases = rollmean(cases, 7, fill=NA)) %>%
filter(!is.na(meancases)) %>%
ggplot(aes(x=date, y=meancases, color=place)) +
geom_line()+
theme_half_open() +
scale_y_continuous(labels = function(x) floor(x/1000)) +
ylab(bquote("\u00D7"*10^3~"number of cases")) +
background_grid(major = "y") +
theme(legend.position = "bottom")
Another way of dealing with jittery curves is to use a sliding window average. This smooths the curves, but partially hides spikes. For example, compare the spike just before (i.e., to the left of) 2021-01, and compare it with the spike without using a sliding window average.
labels <- c("US", "India")
coronavirus %>%
filter(type == "confirmed") %>%
filter(cases > 0) %>%
group_by(country) %>%
# Add US and India to the start of country list, to get the correct legend colors
mutate(country = factor(country, levels = c(labels, sort(setdiff(country, labels))))) %>%
mutate(meancases = rollmean(cases, 7, fill = NA)) %>%
filter(!is.na(meancases)) %>%
ggplot(aes(x=date, y=meancases, color=country)) +
geom_line() +
# highlight countries whose rolling average number of cases exceeds 200,000
# if necessary, connect label to curve with a black line segment
gghighlight(max(meancases) > 200000, label_key=country, label_params = list(segment.color="black")) +
theme_half_open() +
scale_y_continuous(labels = function(x) floor(x/1000)) +
ylab(bquote("\u00D7"*10^3~"number of cases")) +
background_grid(major = "y")
If your line chart contains many curves, it can be useful to highlight the most important ones. This allows you to show all data without overloading the reader.
# create data for line chart
d <- coronavirus %>%
filter(type == "confirmed") %>%
filter(cases > 0) %>%
mutate(place = ifelse(country == "US", "US", "other")) %>%
# order items in legend: first "US", then "other"
mutate(place = factor(place, levels = c("US", "other"))) %>%
group_by(date, place) %>%
summarise(cases = sum(cases), .groups = "drop") %>%
group_by(place) %>%
mutate(meancases = rollmean(cases, 7, fill=NA)) %>%
filter(!is.na(meancases))
# create data for markers
dm <- d %>%
# select dates to place markers
mutate(markers = if_else(mday(date) == 1, meancases, NaN)) %>%
filter(!is.na(markers))
# plot
ggplot(data=d, aes(x=date, y=meancases, color=place)) +
geom_line() +
# plot markers on top of geom_line
geom_point(data=dm, aes(y=markers, shape=place, color=place), size=3) +
gghighlight(label_params = list(segment.color="black"), label_key = place) +
theme_half_open() +
scale_y_continuous(labels = function(x) floor(x/1000)) +
ylab(bquote("\u00D7"*10^3~"number of cases")) +
background_grid(major = "y") +
theme(legend.position = "none")
Line charts can be tricky to read when printed in gray scale or when your reader has a color vision deficiency. A simple way to improve readability is to add markers to your chart.