ggEDA: Visualisations for exploratory data analysis using tiled one-dimensional graphics and parallel coordinate plots

Implementation

ggEDA is implemented as a standard R package and published on CRAN and the R-universe. The interactiveEDA web app was written using the shiny framework⁶ and takes ggEDA as a dependency to separate the user-interface codebase from the underlying business logic, which is easier to test. interactiveEDA is compiled into a purely client-side web-assembly app using shinylive⁷ and hosted as a static web-page on GitHub Pages. Code to produce visualisations is run in the client’s browser instead of a third party server outside the direct control of end-users. The distributed nature of compute also provides scaling benefits compared to traditional server-side shiny apps that quickly slow as concurrent users grow. These security and scalability benefits do come at the cost of slower application startup time.

Operation

The ggEDA R package can be installed from CRAN ( install.packages(“ggEDA”)). It is compatible with Mac OS X, Windows, and all Unix-like operating systems where R (≥3.5.0) can be installed. Package dependencies are described on the ggEDA CRAN listing .

Data can be explored using the interactiveEDA web app in any modern browser that supports WebAssembly. We performed the most extensive testing in Chrome (version 137.0.7151.120), however the app is also compatible with Firefox, Safari and Microsoft Edge.

Creating parallel coordinate plots

PCPs are a well-established EDA visualisation that reveal trends in predominantly quantitative datasets and detect outliers in one or more dimensions. Quantitative features are represented as a series of parallel axes with samples visualised as lines passing through each axis at the point of its value. Correlative relationships are revealed when feature axes are ordered based on line crossing minimisation algorithms or mutual information with a categorical feature. ggEDA can produce interactive PCPs from any dataset with quantitative features using the ggparallel(data) command ( Figure 2).

Figure 2. ggEDA parallel coordinate plots of the dry beans imaging dataset.⁹

A) Visualising 16 morphological features of 13,611 grains from common dry bean species reveals clear correlations amongst size-related attributes (Area, Perimeter and Axis Length). Bombay beans were the largest, most convex variety; B) Highlighting a single subclass simplifies both comparison against the full cohort and identification of within-class outliers. For example, Dermason beans (red) are smaller in size than other varieties. One Dermason bean grain had unusually low roundness, highly atypical for this variety.

PCPs scale well with large datasets but have several limitations. Visualising the relationships between multiple categorical variables is challenging. Missing data is also difficult to meaningfully represent. For this reason, ggEDA introduces a complementary visualisation composed of vertically aligned tile and bar plots.

Creating tiled one-dimensional graphics

For small datasets (n < 1000), ggEDA can represent features as distinct, vertically aligned bar or tile plots, with plot types auto-selected based on whether variables are categorical or numeric ( Figure 3).

Figure 3. ggEDA visualisations of common datasets revealing: A) Petals of the setosa species of iris are drastically smaller than other iris species; B) The majority of individuals who perished during the Titanic disaster were adult males; C) Gentoo penguins from Biscoe Island have shallower bill depths than Chinstrap or Adelie penguins, despite their increased body mass.

Identifying complex multidimensional patterns

To demonstrate the advantages of ggEDA, we created the artificial Lazy Birdwatcher dataset. It describes magpie observations by two birdwatchers, one of whom routinely skips birdwatching on weekends. This introduces a missing data pattern dependent on both the birdwatcher and day of the week. The multidimensional pattern becomes immediately apparent from ggEDA stacked tile plots despite being difficult to detect using one-dimensional EDA tools like skimr, or two-dimensional tools like ggpairs from the GGally package ( Figure 4). Despite being n-dimensional, all PCP plot implementations in R also fail to uncover this trend due to either exclusion of missing data or inability to represent clearly the relationships between categorical features.

Figure 4. Visualisation of the Lazy Birdwatcher dataset using the ggEDA package reveals a pattern of missingness (indicated by exclamation marks) dependent on multiple variables, Birdwatcher and Day (A).

This pattern is difficult to detect using one-dimensional EDA tools like skimr (B) or two-dimensional tools like ggpairs from the GGally package (C).

Exploring datasets using the interactiveEDA web-app

Despite the advancements provided by ggEDA and other tools in the R ecosystem, a key limitation remains: accessibility for non-programmers, particularly when visualising n-dimensional data. All existing R implementations lack graphical user interfaces ( Figure 1). While shiny web apps offer a potential solution, they often require uploading datasets to external servers, raising privacy concerns. To address these limitations, we developed interactiveEDA, a web-assembly compiled client-side web app for secure, interactive data exploration ( Figure 5). Operating entirely in the browser, interactiveEDA ensures data remains on the user’s machine, increasing ease of use without compromising data privacy. interactiveEDA is available at https://github.com/CCICB/interactiveEDA.

Figure 5. Screenshot of interactiveEDA, a web-app providing a graphical user interface for code-free generation of ggEDA visualisations.