Data Visualization

There are several important considerations when choosing colors for data visualization.

The goal in data visualization is to enable the most important data to take center stage. Formatting should support—not detract—from this objective. Chart borders are a usual suspect and prime example of formatting that can divert focus away from the data.

Figure 1 illustrates the use of heavy and light borders for a tabular presentation of data. Formatting takes center stage in the first table due to heavy borders, while formatting is not the focus in the second table given the minimal light gray border.

Two tables have 4 columns and 3 rows. The column headers are dimension, metric 1, metric 2, and metric 3.

The use of a non-zero baseline can exaggerate differences in metrics, resulting in misleading conclusions.

Figure 2 illustrates how minor differences are exaggerated when a non-zero baseline is applied to the y-axis. The average number of hires across these four years is 1000, and the small variation YoY is accurately reflected in the bar chart with a zero baseline.

Two column charts of hires versus years for zero baseline and non-zero baseline. The graph for the baseline has taller columns, and the graph for the non-zero baseline has decreasing columns from the left to the right.

When visualizing data, it is best not to deviate from the way in which the audience will naturally interpret the data. Outlined below are a few important considerations:

Figure 3 shows murders committed using firearms in Florida across time. This visual has an inverted y-axis, which is highly misleading since what appears to be spikes in murders are actually dips. Irrespective of whether this was politically motivated or simply a failed attempt at creativity, this is a deceptive data visualization.

A graph titled gun deaths in Florida plots the number of murders committed using firearms versus years. The curve starts in (1990, 873) and ends in (2010, 721) with fluctuations, In 2005, Florida enacted its Stand Your Ground law.

Preattentive attributes signal where to look and help our audience focus on content we wish to emphasize (Knaflic, 2015).

It is important to remember that data do not always need to be visualized with a chart. Sometimes communicating with text is a sufficient and effective way to convey information. In the case of text, preattentive attributes are often implemented using bold text and/or contrasting colors for keywords and phrases that need to take center stage. Bolded text is generally more effective than italicizing or underlining, as it highlights the chosen elements without unnecessary noise and compromised legibility.

Gray is an important color in the implementation of preattentive attributes as it facilitates a move of less important data and design elements to the background to make room for more predominate colors to highlight the most important information. Figure 4 illustrates the application of preattentive attributes to emphasize the key message in a sentence. This is accomplished by pushing the less important content to the background using gray text and using bold orange text to highlight the labor cost increase.

A paragraph with the following highlighted phrase. Labor costs increasing 30 percent Y o Y.

Preattentive attributes can also be applied to data visualizations to focus the audience’s attention. Figure 5 illustrates the use of preattentive attributes by assigning a different color (orange) to the labor cost for the current year and muting the labor costs associated with prior years using a less predominate gray.

A horizontal bar graph of the year versus labor cost has the following values. The year 2022 has the highest cost of 200, while 2019 has the lowest cost of 100. All values are estimated.

We will begin by building a bar chart that shows the distribution of active employees’ educational backgrounds. To simplify the data structure for data visualization, let us ingest our employees data and create an aggregated cube with counts by ed_field for active employees:

An algorithm with functions for load libraries, load data, and create a data cube for active employees.

A bar chart can be created using the geom_bar() function, and we will start with the ggplot2 library’s design defaults (Fig. 6):

A column chart of n versus e d underscore field for human resources, life sciences, marketing, medical, other, and technical degree with legend. Life sciences have the highest n value, while human resources have the lowest value.

Since the legend is not necessary for interpreting the education field to which each bar corresponds, we will remove it using theme(legend.position = "none"). Eliminating the legend also provides more real estate for the chart, which helps make the x-axis values more legible (Fig. 7).

A column chart of n versus e d underscore field for human resources, life sciences, marketing, medical, other, and technical degree without legend. Life sciences have the highest n value, while human resources have the lowest value.

There is a lot competing for attention due to this vibrant color palette. Let us assume that the objective is to highlight the education field pursued by the largest number of employees. Preattentive attributes lend well to this, and we can assign specific hex color codes to the education categories (Fig. 8).

A column chart of n versus e d underscore field for human resources, life sciences, marketing, medical, other, and technical degree. The highlighted life sciences have the highest n value, while human resources have the lowest value.

Since Life Sciences is the field pursued by the most employees, let us highlight the corresponding bar in blue and move the remaining bars to the background through the assignment of light gray. One method of accomplishing this is via the following explicit category assignments:

Spacing can also be added between the axis titles and labels to improve upon the default formatting and reduce text congestion. The ggplot2::element_text (margin = margin(t = 0, r = 0, b = 0, l = 0)) parameter can be defined for the x and y axes via axis.title.x and axis.title.y, respectively, where:

Let us create a margin of white space above the x-axis title and to the right of the y-axis title by defining the following parameters (Fig. 9):

A column chart of the headcount versus education field for human resources, life sciences, marketing, medical, other, and technical degree. The highlighted life sciences have the highest headcount, while human resources have the lowest count.

We can add a chart title via labs(title = 'This is a chart title.'). This title can be centered via theme(plot.title = ggplot2::element_text (hjust = 0.5), left justified via theme(plot.title = ggplot2::element_ text(hjust = 0), or right justified via theme(plot.title = ggplot2::ele ment_text(hjust = 1).

It is also a best practice to assign a descriptive title to charts to highlight the key message(s) we want to convey to the audience through the visual. For longer titles, the new line character ∖n can be used to break titles into multiple lines (Fig. 10).

A column chart of the headcount versus education field for human resources, life sciences, marketing, medical, other, and technical degree with a title above. The highlighted life sciences have the highest headcount, while human resources have the lowest count.

The default gray background is a distraction from the data we need to take center stage. We can remove this background using panel.background = element_blank() to achieve a cleaner aesthetic and allow the bars in this chart to become more pronounced (Fig. 11).

A column chart of the headcount versus education field for human resources, life sciences, marketing, medical, other, and technical degree without the grid background and a title above. The highlighted life sciences have the highest headcount, while human resources have the lowest count.

Axis ticks add noise to this visual and are not necessary to ascertain to which values along the axes the data align. Axis ticks can be removed for each axis independently with axis.ticks.x = element_blank() for the x-axis and axis.ticks.y = element_blank() for the y-axis (Fig. 12).

A column chart of the headcount versus education field for human resources, life sciences, marketing, medical, other, and technical degree with a title above. The highlighted life sciences have the highest headcount, while human resources have the lowest count.

While the chart and axis titles are important for clarifying what information is represented in the visual, these should not be the focus. Just as we pushed the less important education categories to the background using gray text, we can mute the chart and axis titles with gray text to help draw attention to the data.

We can change the color of the chart title to light gray with plot.title = ggplot2::element_text(colour = "#404040"). The axis titles can be changed to the same color using axis.title.x = ggplot2::element_text (colour = "#404040") for the x-axis and axis.title.y = ggplot2::element_ text(colour = "#404040") for the y-axis (Fig. 13).

A column chart of the headcount versus education field for human resources, life sciences, marketing, medical, other, and technical degree with a title above. The highlighted life sciences have the highest headcount, while human resources have the lowest count.

Flipping the coordinates of axes to convert the default vertical bar chart into a horizontal bar chart allows the audience to more easily scan down the right side of the visual to quickly identify and understand the relative frequencies of education categories (Fig. 14).

A horizontal column chart of the education field versus headcount for the technical degree, other, medical, marketing, life sciences, and human resources with a title above. The highlighted life sciences have the highest headcount, while human resources have the lowest count.

This visual can be further simplified by sorting the bars from highest to lowest value. With sorted bars, the audience can more easily ascertain the relative ranking of each education field.

We can pass reorder(ed_field, n) into the aes() function to sort the education field bars from highest to lowest n-count. If we needed to sort in the opposite direction, reorder(ed_field, -n) will reverse the sort direction (Fig. 15).

A horizontal column chart of the education field versus headcount for the technical degree, other, medical, marketing, life sciences, and human resources sorted from highest to lowest. The highlighted life sciences have the highest headcount, while human resources have the lowest count.

As shown in Fig. 16, the final visual is a marked improvement over the initial design defaults.

A horizontal column chart of the education field versus headcount, and a column chart of n versus e d underscore field for life sciences, medical, marketing, technical degree, other, and human resources. Both have life sciences with the highest headcount and human resources with the lowest count.

A simple cube containing employee counts by department and tenure band will be constructed for demonstrating how to display tabular output:

An algorithm has the functions to append new tenure band column, store aggregate measures to cube, and specify ordered factor.

An algorithm to display output using default settings.

                  
                    ## # A tibble: 14 x 3
                  
                  
                    ## # Groups:   dept [3]
                  
                  
                    ##    dept                   tenure_band     cnt
                  
                  
                    ##    <chr>                  <ord>         <int>
                  
                  
                    ##  1 Human Resources        1-2 Years         7
                  
                  
                    ##  2 Human Resources        3-5 Years        19
                  
                  
                    ##  3 Human Resources        6-10 Years       17
                  
                  
                    ##  4 Human Resources        Over 10 Years     8
                  
                  
                    ##  5 Research & Development 1-2 Years       148
                  
                  
                    ##  6 Research & Development 3-5 Years       262
                  
                  
                    ##  7 Research & Development 6-10 Years      252
                  
                  
                    ##  8 Research & Development Over 10 Years   148
                  
                  
                    ##  9 Research & Development Under 1 Year     18
                  
                  
                    ## 10 Sales                  1-2 Years        57
                  
                  
                    ## 11 Sales                  3-5 Years        93
                  
                  
                    ## 12 Sales                  6-10 Years      124
                  
                  
                    ## 13 Sales                  Over 10 Years    70
                  
                  
                    ## 14 Sales                  Under 1 Year     10
                  
                

While these data could easily be copied and pasted into presentation software for formatting, additional libraries exist in R for formatting tables. For example, R Markdown scripts can leverage the DT package to provide filtering, pagination, sorting, search, and other interactive features for HTML output. Field names can also be changed to proper case via the dplyr::rename() function (Fig. 17).

A screenshot of 10 tabulated entries. The table has three columns, namely, department, tenure, and count.

An algorithm for the assign proper case field names function.

The geom_tile() function can be used to build a heatmap with ggplot2. The range of colors can be specified via the scale_fill_continuous(low = 'minimum value color', high = 'maximum value color') function since the fill variable defined by aes(fill = cnt) is measured on a continuous scale (Fig. 18).

A heatmap of department versus tenure plots employee count ranging from 50 to 250. The text on the top is as follows. A large portion of our workforce has between 3 and 10 years of company tenure and works within research and development.

This heatmap is excellent for focusing attention on the department + tenure segments with highest and lowest employee counts. However, if the specific employee counts are required, complimenting this heatmap with a basic table of metrics is a good option to avoid cluttering the heatmap with 3 ∗ 5 = 15 additional numbers.

Let us build a scatterplot to visualize the relationship between work_exp and ytd_sales, with a goal of focusing the audience’s attention on salespeople whose ytd_sales meets or exceeds the full year sales quota of 150, 000 USD. Let us first subset employees data to active salespeople and append a flag to indicate sales quota attainment for use in shading data points in the scatterplot.

An algorithm for the subset d f to active sales employees and set quote attainment flag for data viz coloring functions.

The geom_point() function is used to create a scatterplot. The scale_y_ continuous() function can be used in conjunction with the scales library to override the default scale for the y-axis (scientific notation) with more intuitive values.

While the basic scatterplot in Fig. 19 is effective in visualizing the relationship between work_exp and ytd_sales, additional design elements are needed to highlight the data points that meet or exceed the full year sales quota of 150, 000 USD.

A scatterplot of Y T D sales versus work experience has denser values on the left and scattered values on the right. The text at the top reads the relationship between work experience and Y T D sales is positive.

The color argument for the geom_point() function is used to apply preattentive attributes to this visual by specifying the field used for conditional data point shading. We can define the color for each of the field’s values using the scale_color_manual() function. Additionally, the geom_hline() function is used to add a dotted horizontal line at a specified position on the y-axis (at 150, 000 USD), and annotation is added at a specified pair of x and y coordinates using the annotate() function (Fig. 20).

A scatterplot of Y T D sales versus work experience has denser values on the left and scattered values on the right, the sales quote is mentioned as equal to 150 thousand dollars. The text at the top reads employees with more work experience have already met the full-year sales quota.

When visualizing trended data, it is important to avoid cumulative trends because they suggest an upward trajectory (positive slope) even when the corresponding non-cumulative metrics indicate a declining trend.

Consider the following data frame with decreasing hire counts by year:

An algorithm for the print hires d f function.

                  
                    ##   year hires cum_hires
                  
                  
                    ## 4 2019  1020       980
                  
                  
                    ## 3 2020  1010      1970
                  
                  
                    ## 2 2021   990      2980
                  
                  
                    ## 1 2022   980      4000
                  
                

Figure 21 juxtaposes a trended line chart with hires by year (left) against a cumulative version of the same (right). Since hires at each year are additive in the cumulative line chart, the slope is positive and misleading.

Two graphs of hires versus years for trended hires and cumulating trended hires have a decreasing line and an upward-sloping line, respectively.

An algorithm with the functions to set seed for reproducibility and create data.

Line graphs can be constructed using the geom_line() function in ggplot2. Figure 22 shows a single series line graph built in ggplot2.

A graph of attrition rate versus month plots a concave up, increasing curve. The text on the top reads engineering attrition has increased over the last 24 months, largely due to an exponential spike in the last year.

Common applications for slopegraphs in people analytics include survey variable score changes between two time periods, pre/post changes to outcome variables in an experimental context, and various metrics (e.g., headcount, TTM attrition) for which changes need to be evaluated MoM, QoQ, or YoY when data points between the start and end points are unimportant.

The data structure needed to support a slopegraph is consistent with a line graph. To illustrate how to construct a slopegraph in R, a data frame will be constructed that holds engagement scores for two points in time for both a treatment and a control group.

An algorithm with the function to build data frame with Y o Y headcount metrics by the department.

In ggplot2, a slopegraph is simply a line graph with only two values on the x-axis and some additional formatting including annotations and y-axis removal. Figure 25 is a slopegraph comparing the engagement score changes for treatment and control groups over an observation period, with preattentive attributes applied to focus attention on the treatment group.

A graph illustrates the observation period from time 1 to time 2 with a 16% increment for the control group and a 55% increment for the treatment group. The text at the top reads a more favorable change in engagement was observed for the treatment group relative to the control group.

Bar charts are used to display categorical data. Four common types of bar charts are vertical, horizontal, stacked, and bidirectional.

An algorithm with functions for data preparation, set seed for reproducibility, generate favorability distributions, average top box favorable score, and build data frame with Y o Y headcount metrics by the department, among others.

The geom_bar() function can be leveraged to construct a bar chart using ggplot2. The geom_hline() and annotate() functions can be added to include a reference line with the average top box score. Preattentive attributes can also be applied by setting the derived top_score variable as the fill parameter in the aes() function (Fig. 26).

A bar chart of engagement score versus departments for people, finance, marketing, product, legal, engineering, and sales. People have the highest highlighted score, while sales have the lowest score. The average is given as 33.

An algorithm with the functions to set seed for reproducibility, build data frame with hire and term metrics by department, and append transformed count column to support bidirectional bar charts.

Using the cnt_trans field containing both negative and positive integers, we can visualize net changes in departmental headcount (Fig. 29).

A stacked bar chart of department versus the headcount changes plots terms and hires for sales, product, engineering, finance, marketing legal, and people. Terms have eclipsed hires in sales, product, and engineering over the trailing 12-month period.

In a people analytics context, we may wish to highlight differences between regrettable (bad) and non-regrettable (good) turnover trends relative to total voluntary turnover rates. It is usually more difficult to compare the magnitude of regrettable and non-regrettable rates across time using a stacked bar chart, and a combination chart is often a more intuitive method of presenting this information.

As illustrated in Fig. 30, we can leverage a two-series line chart for monthly regrettable and non-regrettable rates relative to the total voluntary turnover rate visualized with a light gray vertical bar chart in the background.

A graph of attrition rate versus month plots fluctuating curves for regrettable and not regrettable. The text at the top reads the largest share of monthly-voluntary attrition over the 24-month period has consistently been regrettable terminations.

Building a waterfall chart requires a bit more data prep relative to the data structure requirements for a stacked bar chart. Unlike a bar chart, a waterfall chart needs a start and an end value for each event type so that each bar beyond the first begins where the previous bar ends.

An algorithm with the functions to generate headcount data with i d field to define bar order, define start and end values to support waterfall chart, and swap start or end values for the last record of the ending H C.

With data properly structured, we can use the geom_rect() function to build the waterfall chart. Strategies such as labeling only the Beginning HC and Ending HC reduces clutter and lends to a cleaner design aesthetic (Fig. 31).

A waterfall chart of headcount versus event for starting H C, hires, transfers in, transfers out, exits, and ending H C. The text at the top reads hires and inbound transfers have outpaced outbound transfers and exits in engineering, leading to a net Y T D increase in headcount.

Waffle charts, also known as square area charts, are well-suited for illustrating parts of a whole.

A common application in people analytics is illustrating candidate movement through the recruiting funnel. It is often helpful to visualize a normalized applicant pool (e.g., per 100 applicants) to identify where bottlenecks exist across funnel stages and how pass-through rates are compared across business areas.

Waffle charts require some data prep to visualize:

An algorithm with the functions to create d f with T A funnel metrics, set depth of waffle chart, expand the applicant counts into a vector of stage labels, and integrate stages and fill any extra tiles with N A, among others.

With this data structure, a waffle chart can be produced using the geom_tile() function from ggplot2.

The waffle chart in Fig. 32 illustrates how candidates for a set of filled requisitions move from application through the phone screen, interview, offer extend, and offer acceptance stages of the recruiting lifecycle. Based on this chart, 60% of applicants are rejected without a conversation, 40% receive a phone screen, 20% land an interview, 10% receive an offer, and 4% accept an offer.

A waffle chart in a 10 by 10 layout highlights the grids for apply, phone screen, interview, offer extend, and offer accept in different color shades. The text on the top reads we select 10 in every 100 applicants. Our offer acceptance rate, O A R is 40%.

Sankey diagrams, or alternatives such as chord diagrams and alluvial plots, are flow diagrams that are particularly effective in depicting a many-to-many relationship between two domains.

Sankey diagrams have many applications in people analytics. For example, sankeys may be used to understand internal transfers over a period of time (i.e., inflows and outflows among departments) or recruiting sources by which employees in various departments have been hired. To demonstrate how to implement a sankey diagram, let us use the sankeyNetwork() function from the networkD3 library to visualize employee transfers between departments. D3 is an advanced JavaScript framework for creating interactive visualizations, and the networkD3 library provides an easy interface for constructing sankey diagrams with interactive components. Building a sankey diagram using networkD3 requires data to be structured within two data frames:

An algorithm with the functions to set seed for reproducibility, create nodes d f, create links d f, and append employee transfer counts per department pair to links.

With data properly structured within the nodes and links data frames, these data can be passed into the sankeyNetwork() function to construct the sankey diagram. As illustrated in Fig. 33, hovering over a connection displays the count of employee transfers between the corresponding departments (n = 222 moves within Product).

A Sankey diagram with a dense interconnected web pattern between the nodes labeled engineering, finance, legal, marketing, people, product, and sales.

Though it can quickly become noisy as the number of nodes increases, coloring connections based on the source departments from which each group flows can be helpful in tracing the employee flow to the various destination departments.

A department variable needs to be defined and added to the links data frame in order to color connections based on the source department (transfers out). The NodeGroup and LinkGroup parameters in the call to the sankeyNetwork() function define the column in the nodes and links data frames, respectively, that specify the group values by which the nodes and links should be colored.

An algorithm with the append source department variables for colored connections function.

We can add another grouping variable to support coloring intradepartmental (within) moves differently from interdepartmental (outside) moves (Fig. 34).

A color gradient Sankey diagram with a dense interconnected web pattern between the nodes labeled engineering, finance, legal, marketing, people, product, and sales.

An algorithm with the append within slash outside department transfer variable for colored connections function.

If the focus is on a single department’s transfers, preattentive attributes can be applied to help focus the audience’s attention and move the less important information to the background to reduce clutter (Figs. 35 and 36).

A color gradient Sankey diagram with a dense interconnected web pattern between the nodes labeled engineering, finance, legal, marketing, people, product, and sales.

A Sankey diagram with a dense interconnected web pattern between the nodes labeled engineering, finance, legal, marketing, people, product, and sales. The connections for the product are highlighted.

An algorithm with the append dichotomous product slash other indicator to links and nodes data frames function.

Figure 37 shows a donut chart—which is simply a pie chart with the center cut out—that is intended to show Kane Williamson’s outsized contribution to New Zealand’s cricket runs. However, top scorers from other countries are also included in this visual, so the metrics do not sum to 100%. Pie and donut charts are not appropriate for these data.

A donut chart titled the world cup's big guns, percentage of team's runs scored by top scorer includes some of the following values. Williamson, 30.23. Rohit, 29.05. Shakib, 28.25, Warner, 25.02. Rahmat Shah, 14.8, among others.

There is no specific geom() function for building pie charts using ggplot2. Therefore, we need to create a bar chart and make it circular by adding a coord_polar("y", start = 0) transformation. The coord_polar() transformation complicates the positioning of labels, but we can add a position = position_stack(vjust = 0.5)) argument within the geom_text() function to easily achieve metric centering within the respective category. We can also easily improve the aesthetics with the theme_void() function, which removes the default background, grid, and labels (Fig. 38).

A pie chart has the following values. Female, 40.6%. Male, 59.4%. The text at the top reads nearly 60 percent of active employees identify as male.

An algorithm with the function to create a data cube for gender representation among active employees.

An element of uncertainty is inherent in analytics, which is why this book avoids the use of categorical terms such as always and never. However, an exception will be made for this section. 3D visualizations are never appropriate. 3D visuals often misrepresent the data they are intended to visualize and are unnecessarily difficult to interpret.

Figure 39 shows the distribution of headcount across locations using a 3D pie chart. Based on this visual, it may not be apparent that Los Angeles (25%) is considerably larger than New York (19%). The misleading perspective of this tilted visual causes locations on the back side of the pie chart to appear smaller relative to locations on the front side.

A 3 D pie chart titled headcount by location has the largest count for New York, followed by Chicago, Los Angeles, London, and Tokyo.

A 3D bar chart is not an improvement over a 3D pie chart. As illustrated in Fig. 40, bars do not appear to align with the actual corresponding values (e.g., Los Angeles =  25%). This is because software determines the height of bars in a 3D bar chart using an invisible tangent plane to intersect the bars at the correct y-axis location.

A 3 D bar chart of headcount by location plots data in percentage versus Los Angeles, London, New York, Chicago, and Tokyo. Los Angeles has the highest count, while Tokyo has the lowest count.

As shown in Fig. 41, by leveraging a horizontal bar chart with labeled bars and preattentive attributes, it becomes easy to understand relative headcount across locations.

A bar chart of headcount by location with the following values. Los Angeles, 25%. London, 23%. New York, 19%. Chicago, 17%. Tokyo, 16%.