corrViz

Introduction

corrViz is an R package designed to help users visualize correlations between variables in their datasets. With the ever-growing size and complexity of datasets, it is crucial to have a clear understanding of the relationships between variables. This package aims to provide an easy and effective way to explore and visualize these correlations, making it easier to interpret and communicate results.

This vignette will introduce you to the main features and functions of the corrViz package, guiding you through the process of creating visually appealing and informative correlation plots. By the end of this tutorial, you should be able to use corrViz to create various types of correlation visualizations, customize their appearance, and gain insights into the relationships within your dataset.

Download the development version from GitHub with:

# install.packages("devtools")
devtools::install_github("AlanInglis/corrViz")

You can then load the package with:

library(corrViz)

Methods

The corrViz package offers several visualization methods to help you explore and communicate correlations effectively, with an emphasis on interactivity. The only required inputs for all corrViz functions are either a dataset or a square matrix of correlations to be visualized. Each visualization has multiple customization options, which are outlined in this document. The plotly package1 is used to create interactive graphics.

To begin we will create a correlation matrix using the mtcars data set.

cm <- cor(mtcars)

Heat map

The following code produces an interactive heatmap displaying variable correlations, which can be seen in Figure 1. The additional arguments here are:

Hovering the mouse over a cell in the heatmap will display a box containing the variable pair names and their associated correlation value. Dragging the mouse will zoom in on the selected area and double-clicking the plot will reset the zoom.

corrHeatmap(mat = cm,
            display = 'all',
            reorder = TRUE,
            pal = colorRampPalette(c("darkblue", 'white', 'darkred'))(100))
Figure 1: Interactive heatmap displaying all correlations.

In Figure 1 we set display = "all" which shows the full heatmap. To display only the upper or lower portion of the heatmap we use the appropriate argument in display, as shown in Figure 2 (which shows the lower section).

corrHeatmap(mat = cm,
            display = 'lower',
            reorder = FALSE)
Figure 2: Interactive heatmap displaying correlations from the lower section of the heatmap.

Network

For build our network plots we use the visNetwork package2. the additional arguments for producing a network plot are:

corrNetwork(mat = cm,
            threshold = 0,
            layout = "layout_nicely",
            width = "100%",
            height = "400px",
            physics = FALSE)
Figure 3: Interactive network plot displaying correlations.

Figure 3 shows a network in which every node corresponds to a variable present in the data, and each connecting edge denotes the correlation between them. The edges are color-coded to signify the correlation, with red indicating a positive correlation and blue representing a negative one. The thickness of each edge corresponds to the magnitude of the correlation between two variables. Hovering the mouse over an edge will display a box containing the correlation value for that edge.

When clicking on a node, the selected node will change colour and only nodes that connected to the selected node are highlighted (a better example of this can be seen in Figure 4). Additionally, the top right of the plot has a drop-down box which can be used to select a particular node/variable. DOing so will have the same effect as clicking on the node. The nodes themselves can be re-positioned by dragging them (only works when physics = FALSE). The bottom left of the plot contains controls to zoom in or out, whereas the bottom right has controls to pan the image.

In Figure 4, below, we set the threshold argument to equal 0.8. This will filter any correlations from our visualization with an absolute value lower than 0.8. Additionally, in Figure 4, physics = TRUE. By enabling physics, each node is repelled from each other and the edges act like springs. When physics = TRUE, the nodes can still be dragged, however as each node repels each other with equal force, the display will try to push each node away from each other.

corrNetwork(mat = cm,
            threshold = 0.8,
            physics = TRUE)
Figure 4: Interactive network plot displaying variables with an absolute correlation value above 0.8.

Solar plot

In a solar system correlation plot, the dependent (sun) variable of interest is positioned at the center, represented as the sun. The explanatory (planet) variables are depicted as planets orbiting around the sun, with their distance from the sun corresponding to the absolute value of their correlation with the dependent variable. Therefore, the greater the distance of a planet from the sun, the weaker the correlation between the explanatory variable and the dependent variable. # Solar plot animation

In Figure 5 we show an animated version which displays the planet variables orbiting the sun. The additional arguments are:

animSolar(mat = cm,
          sun = 'mpg',
          export = FALSE,
          num_frames = 100,
          path = NULL,
          gif_name = "solar_system.gif",
          fps = 60)
Figure 5: Animated solar system correlation plot with the variable mpg as the sun.

The num_frames argument is used to select the number of frames. Setting this to a low value will produce the plot quicker, however having a low number of frames will result in the “planets” jumping as the frames transition. Additionally, a low values of num_frames will affect the orbit of the animation when setting export = FALSE. This differs from the fps argument which sets the number of frames to play per second for use when exporting a gif. When export = FALSE (as is the case in figure 6), a ‘play’ button is provided to start the animation. Hovering over a planet will display the correlation value.

Bar plot

Figure 6 shows an interactive barplot displaying correlation values. The red bars extending right represent the positive correlations whereas the blue bars extending left represent the negative correlations. The additional arguements here are: * interactive: If TRUE then an interactive version of the barplot is displayed. * pal: The colour palette to use for displaying values.

corrBarplot(mat = cm,
            interactive = TRUE,
            pal = colorRampPalette(c("cornflowerblue", 'white', 'tomato'))(100))
Figure 6: Bar plot displaying correlation values.

Hovering the mouse over a bar will display the name of the variable pair and their associated correlation.

Bubble

Figure 7 is similar to the heatmap shown in Figures 1 and 2, except this time each variable pair is represented as a circle, with the size corresponding the the magnitude of the correlation. Hovering over a bubble will display the variable pair names and their associated correlation. The additional arguments for this function are:

corrBubble(mat = cm,
           display = 'all',
           pal = colorRampPalette(c("cornflowerblue", "white", "tomato"))(100))
Figure 7: Bubble plot displaying correlation values.

As with the heatmap, we can also display a portion of the plot using the display argument, as shown in Figure 8.

corrBubble(mat = cm,
           display = 'upper')
Figure 8: Bubble plot displaying correlation values from the upper portion of the correlation matrix.

Scatterplot matrix pairs plot

Our next plot is a scatterplot matrix (or pairs plot) displaying correlation values in the upper triangle of the plot, with the stars representing the correlation significance. The diagonal contains density plots of the data. The lower triangle displays a scatterplot of the data values with a smooth-line (which is coloured according the the correlation value for the variable pair). The additional arguments are:

Figure 9 shows an interactive scatterplot for the first 5 variables found in the mtcars dataset. In this case, the scatterplot points are coloured by the variable cyl.


corrPairs(data = mtcars[1:5],
         interactive = TRUE,
         col_by = "cyl")
Figure 9: Correlation pairs plot for the first five variables from the mtcars dataset.

Sankey plot

An alternative method for viewing correlation is by using a Sankey plot. Sankey plots are primarily used to represent the flow of information between nodes in a network. These diagrams consist of interconnected pathways, with their width proportionally representing the quantity being transferred (in this case, the absolute value of the correlation). Figure 10 provides an example. In this case the correlations are filtered to show correlations with an absolute value above 0.6. The additional arguments are:

corrSankey(mat = cm,
           threshold = 0.6,
           colour = FALSE)
Figure 10: Sankey plot displaying correlation values with a magnitude above 0.6.

Setting colour = TRUE colours the links based on their correlation. In figure 11, the threshold argument is set to equal 0.8.

corrSankey(mat = cm,
           threshold = 0.8,
           colour = TRUE)
Figure 11: Coloured Sankey plot displaying correlation values with a magnitude above 0.8.

Shiny Plot

We also provide an option to create an interactive Shiny plot. This plot is a useful tool for quickly checking the correlation between two variables. In Figure 12 below we show a screenshot of the Shiny app in use. The additional arguments are:

corrShiny(data = mtcars,
          x_var = "wt",
          y_var = "mpg",
          color_var = "cyl",
          size_var = "hp")
Figure 12: Shiny correlation explorer.

In the code above, although we chose with variables to display, colour and size by, all of these options options can be changrd within the Shiny app. As can be seen in Figure 12, we provide the options to choose both the X and Y variable to display in the plotting window. The correlation value is displayed in the Figure caption or can be displayed in the bottom right of the plot by checking the show correlation coefficient on plot box. The correlation method can be changed by clicking on the appropriate drop-down menu and we provide three correlation method options. They are; Pearson, Spearman, or Kendall.

A line of best fit (with or without confidence intervals) can be overlayed on top off the data and provide two options to fit the line, i.e., \(lm\) of \(loess\). To aid in interpretability, we also provide options to colour the points in the plot by a selected variable, to size the points by a selected variable, or to show observation labels on the plot.

Static Plots

In addition to interactive plots, corrViz also provides numerous static plots. In Figures 13 to 16 we show a selection of options to display correlations on a grid. The plot types can be any of square, circle, text, or pie, which can be set via the type argument. If the type argument is set to either square or circle, then the shapes are sized, with the size representing the absolute value of the correlation. The additional arguments are:

In Figure 13, we set type = 'square'

corrGrid(mat = cm[1:7,1:7],
         type = 'square')
Figure 13: Static heatmap correlation plot.

In Figure 14, we set type = 'circle'.

corrGrid(mat = cm[1:7,1:7],
         type = 'circle')
Figure 14: Static circle correlation plot.

In Figure 15, we set type = 'text'. The text

corrGrid(mat = cm[1:7,1:7],
         type = 'text')
Figure 15: Static text correlation plot.

Correlation Pie Plot

In Figure 16, we set type = 'pie'. This plot shows pie charts where the filled in percentage represents the correlation value.

corrGrid(mat = cm[1:7,1:7],
         type = 'pie')
Figure 16: Static pie-chart correlation plot.

Static Solar Plot

Figure 17 shows a static version of the solar system correlation plot with mpg selected as the dependent variable via the sun argument. Planet variables are coloured either red (positive correlation) or blue (negative correlation).

corrSolar(mat = cm,
          sun = 'mpg')
Figure 17: Solar system correlation plot with the variable mpg as the sun.

Chord plot

Figure 18 shows a chord plot displaying correlations. This type of plot can quickly become overloaded with information, so it is recommended to only use this plot when thresholding correlation values. In Figure 18, threshold = 0.8, which filters any correlations from the visualization an absolute value lower than selected value. The width of the chords corresponds to the magnitude of the correlation with their colours indicating positive (red) or negative correlations (blue). The additional arguments are:

corrChord(mat = cm,
          circle = FALSE, 
          threshold = 0.8)
Figure 18: Chord plot displaying correlation values with a magnitude above 0.8.

Figure 19, displays the same information as Figure 18, but uses a circular layout, via the circle argument.

corrChord(mat = cm, 
          circle = TRUE, 
          threshold = 0.8)
Figure 19: Circular chord plot displaying correlation values with a magnitude above 0.8.

Circular correlation plot

Figure 20 shows a circular correlation plot. This is similar to Figure 19 but places the variable names on the outside of a neater circle. As with Figures 18 and 19, the width of the chords in Figure 20 corresponds to the magnitude of the correlation with their colours indicating positive (red) or negative correlations (blue). The additional arguments are: * ticks: If TRUE, then axis ticks are displayed on the circle. * threshold: Filter correlations with an absolute value lower than selected value.

corrCircle(mat = cm, 
           threshold = 0,
           ticks = TRUE)
Figure 20: Circular correlation plot displaying correlation values.

In Figure 21, we filter the visualization to display correlations with an absolute value greater than 0.8. In this case the axis ticks are removed.

corrCircle(mat = cm, 
           ticks = FALSE, 
           threshold = 0.8)
Figure 21: Circular correlation plot displaying correlation values with a magnitude above 0.8.

Coverting a correlation matrix to data frame

Often it can be useful to convert a matrix to a long-format data frame. The martrix2long function provides just that. The only input is a matrix. Below a demonstartion is shown for the first four colummns of the mtcars data.

corr_matrix <- cor(mtcars[1:4])
long_data_frame <- matrix2long(mat = corr_matrix)
long_data_frame
#>    row_name col_name      value id
#> 1       mpg      mpg  1.0000000  1
#> 2       cyl      mpg -0.8521620  1
#> 3      disp      mpg -0.8475514  1
#> 4        hp      mpg -0.7761684  1
#> 5       mpg      cyl -0.8521620  2
#> 6       cyl      cyl  1.0000000  2
#> 7      disp      cyl  0.9020329  2
#> 8        hp      cyl  0.8324475  2
#> 9       mpg     disp -0.8475514  3
#> 10      cyl     disp  0.9020329  3
#> 11     disp     disp  1.0000000  3
#> 12       hp     disp  0.7909486  3
#> 13      mpg       hp -0.7761684  4
#> 14      cyl       hp  0.8324475  4
#> 15     disp       hp  0.7909486  4
#> 16       hp       hp  1.0000000  4

Conclusion

The corrViz package provides a convenient and powerful way to visualize correlations between variables in your dataset. With its variety of visualization methods and customization options, you can gain valuable insights into the relationships within your data and communicate these findings effectively.

  1. C. Sievert. Interactive Web-Based Data Visualization with R, plotly, and shiny. Chapman and Hall/CRC Florida, 2020.↩︎

  2. Almende B.V. and Contributors, Thieurmel B (2022). visNetwork: Network Visualization using ‘vis.js’ Library. R package version 2.1.2, https://CRAN.R-project.org/package=visNetwork↩︎

  3. Csardi G, Nepusz T: The igraph software package for complex network research, InterJournal, Complex Systems 1695. 2006. https://igraph.org↩︎