High-dimensional visualization methods for data envelopment
analysis (DEA), gathering in one place techniques that have appeared in the
literature but remained scattered and largely unimplemented: cross-efficiency
matrix unfolding, the Porembski network with lambda edges,
principal component analysis biplots, multidimensional-scaling
colour-plots, self-organizing maps, the Costa bi-dimensional efficient
frontier, parallel coordinates, radar charts, panel-data trajectory biplots,
peer and reference networks, and a set of descriptive plots. The package is
built around a single validated dea_data() object and uses the 'Benchmarking'
package as its DEA engine. The implemented methods draw on a body of
literature; representative references include Doyle and Green (1994)
High-dimensional visualization methods for data envelopment analysis (DEA), gathered into a single R package. It brings together techniques that have appeared in the literature but remained scattered and largely unimplemented: cross-efficiency matrix unfolding, the Porembski reference-set graph, PCA biplots, multidimensional-scaling colour-plots, self-organizing maps, the Costa bi-dimensional efficient frontier, parallel coordinates, radar charts, panel-data trajectory biplots, peer and reference networks, and a set of descriptive plots.
The methods follow the author's PhD thesis (Ashkiani, 2019).
Everything is built around one validated object, dea_data, which every
function accepts (plain data frames are coerced automatically). Functions follow
a consistent naming convention plot_<source>_<method>:
compute_* — analysis, returns values (efficiency scores, cross-efficiency
matrices, multiplier weights, a fitted SOM).plot_io_* — plots built from inputs/outputs (the io).plot_cem_* — plots of the cross-efficiency matrix (the cem).plot_efficiency_* — plots of the efficiency scores.plot_panel_io_* — plots of panel (long-format) input/output data.Most plotting functions share the same four arguments:
interactive — FALSE (default) returns a static ggplot2 object;
TRUE returns an interactive plotly object.labels — TRUE (default) shows the identifying text labels for that plot
(DMU names on points and networks, DMU ticks on heatmaps, variable names on
distributions and radar, and so on).title — an optional plot title.... — passed through to the underlying geom (or to the relevant compute_*
function for the plots that take raw data).A shared colour scheme is used throughout: the colour-blind-safe Okabe–Ito palette for categories and viridis for continuous scales.
The Benchmarking package is the DEA engine. The modelling and plotting
packages are in Suggests; each function checks for the ones it needs and
errors informatively if they are missing.
# install.packages("devtools")
devtools::install_local("deaviz") # from the package directory
# install the suggested engines you intend to use, e.g.:
install.packages(c("Benchmarking", "ggplot2", "plotly", "lpSolve",
"smacof", "kohonen", "MASS", "ggrepel"))
library(deaviz)
# Build the data object (the example data is not i_/o_ prefixed,
# so name the inputs and outputs explicitly)
d <- dea_data(
chinese_cities,
inputs = c("industrial_labour_force", "working_funds", "investments"),
outputs = c("gross_industrial_output", "profit_and_tax", "retail_sales"),
id = "DMU"
)
# Analysis
eff <- compute_efficiency(d, rts = "crs") # Farrell efficiency
ce <- compute_cross_efficiency(d) # cross-efficiency matrix
# Visualization (static ggplot2 by default)
plot_efficiency_distributions(d)
plot_io_pca_biplot(d)
plot_cem_heatmap(ce)
plot_cem_unfolding(ce)
plot_io_costa_frontier(d)
plot_io_peer_network(d, size_by_peers = TRUE)
# The same plot, interactive (plotly), with labels turned off
plot_io_pca_biplot(d, interactive = TRUE, labels = FALSE)
With your own data you can skip the explicit inputs/outputs by prefixing
columns i_ (inputs) and o_ (outputs); dea_data() then recognises them
automatically.
| Area | Functions |
|---|---|
| Data object | dea_data(), as_dea_data() |
| Efficiency analysis | compute_efficiency(), compute_multiplier_weights() |
| Cross-efficiency analysis | compute_cross_efficiency(), compute_cross_efficiency_weights(), standardize_weights() |
| SOM analysis | compute_som() |
| Descriptive (inputs/outputs) | plot_io_distributions(), plot_io_scatter(), plot_io_heatmap(), plot_io_parcoo(), plot_io_radar() |
| Efficiency plots | plot_efficiency_distributions(), plot_io_efficients() |
| Projections & frontier | plot_io_pca_biplot(), plot_io_mds(), plot_io_costa_frontier(), plot_io_3dscatter() |
| Reference networks | plot_io_lambda_network(), plot_io_peer_network() |
| Panel data | plot_panel_io_biplot() |
| Self-organizing maps | plot_io_som(), plot_io_som_components() |
| Cross-efficiency matrix | plot_cem_heatmap(), plot_cem_unfolding(), plot_cem_weights_heatmap() |
| Data | chinese_cities |
If you use this package, please cite the thesis it is based on:
Ashkiani, S. (2019). Four Essays on Data Visualization and Anomaly Detection of Data Envelopment Analysis Problems [PhD thesis, Universitat Autonoma de Barcelona]. https://ddd.uab.cat/record/240333
See citation("deaviz") for the BibTeX entry.
This package implements my DEA-Viz methods. An AI coding assistant (Claude, Anthropic) was used to help implement and refine the R code from my specifications and prototype functions; I designed the methods, and reviewed, tested, and validated all code, taking full responsibility for the software.
AGPL-3.