R-Forge Logo


CRAN Github license Downloads

This package provides the implementation of the radar-boxplot, a chart created and developed by the author.


The package is available in the CRAN repository.



There are two variants Default and Formula that accepts two different sets of arguments as input.



Additional optional arguments


The radar-boxplot merges the concepts of both radar chart and the boxplot, allowing to compare multivariate data for multiple classes/clusters at the same time. It provides an intuitive understanding over the data by creating radar polygons which can be compared in terms of shape and thickness, giving a meaningful insight towards identifying high inner variation and similar classes/clusters.

By interpreting the radar-boxplot, it is possible to predict classification confusion over classes and understand why and what could be done to achieve better results.

How it works

The plot consists of a polar plot with radius 1. The circunference is divided into n axes, where n is the number of attributes which will be represented, one for each axis. The attributes are normalized to range from 0.1-1.0 to avoid overlapping low values in the center of the plot. The normalization is done through normalized = (X - min) / (max - min), then it is transformed to 0.1-1.0 range through rescaled_norm = (normalized * 0.9) + 0.1.

The radar-boxplot draws two different colored regions representing the same a boxplot would, but for multiple attributes at once. The inner red region represents the 25-75% percentiles of each attribute, while the blue area represents the total range, excluding the outliers as defined by Tukey (1977). Outliers appears as points, just like the classic boxplot and are defined with the formulas:

IQR = Q3 - Q1
LOWER_OUTLIER = Q1 - (1.5 x IQR)
UPPER_OUTLIER = Q3 + (1.5 x IQR)

The following shows an example of the radar-boxplot over the Wine Quality Dataset (Cortez et al., 2009).

Radar-boxplot example with red wine quality dataset

As the rating gets higher there are two different things happening. First, there is a shift on the overall shape, mainly getting a more defined "pointy" shape towards citric acid and alcohol, while concave for volatile acidity. The second interesting fact is that the inner variation appears to reduce, suggesting that top quality wines must conform to a stricter set of parameters, while intermediante ones can have a mixture of poor properties along with high quality ones compensating each other. I could also propose a cluster analysis within ratings 5 and 6 (because of the high inner variation) to try to understand if there are multiple patterns within them, which could reveal different sets of intermediate wines.

The radar-boxplot is best suited for when there are more than 4 relevant variables for your clustering/classification task, because it gives the possibility to represent higher dimensionality while still being readable.



# Regular
radarBoxplot(quality ~ ., winequality_red)

# Orange and green pattern with white median
orange = "#FFA500CC"
green = rgb(0, .7, 0, 0.6)
radarBoxplot(quality ~ ., winequality_red,
       use.ggplot2=FALSE, medianLine=list(col="white"),

# Plot in 2 rows and 3 columns
# change columns order (clockwise)
radarBoxplot(quality ~ volatile.acidity + citric.acid +
       residual.sugar + fixed.acidity + chlorides +
       free.sulfur.dioxide + total.sulfur.dioxide +
       density + pH + sulphates + alcohol,
       data = winequality_red,


Thanks Dr. Michael Friendly for your great suggestions for improving this package, I'm still working on those. Also, I'd like to thank Dr. Peter Rousseeuw for his valuable feedback.