dgraphs constructs and analyzes graphs derived from numerical observations.
It includes mutual and shared-neighbor graphs, intersection and geodesic
nearest-neighbor graphs, radius and adaptive-radius graphs, and
minimum-spanning-tree completion. Utilities for conversion, pruning,
diagnostics, spectral embedding, endpoints, and paths are also provided.
This page and the website document development version 0.3.0.9000. CRAN currently supplies 0.2.0 (checked 15 September 2026), which has an older public API. Changing the repository version does not publish a CRAN release.
For all examples on this page, install the development package, including its offline guides. A C++17 toolchain and Pandoc are required for this source installation (Rtools on Windows; Xcode command-line tools on macOS):
install.packages(c("remotes", "knitr", "rmarkdown"))
remotes::install_github("pgajer/dgraphs", dependencies = NA,
build_vignettes = TRUE)
library(dgraphs)
packageVersion("dgraphs") # 0.3.0.9000
The guides include static figures without optional viewers. For the 3D
quadratic-form galleries, install htmlwidgets and rgl, then install the
pinned viewer before building dgraphs:
install.packages(c("htmlwidgets", "rgl"))
remotes::install_github("pgajer/ivue@a952ab6816636a19f2038ba0709181608b71770d",
dependencies = NA)
To stay on CRAN 0.2.0, use this separate, release-compatible example:
install.packages("dgraphs")
library(dgraphs)
set.seed(1)
x <- matrix(rnorm(80), ncol = 2)
graph <- create.mknn.graph(x, k = 4)
print(graph)
The remaining examples use the development installation above. Breaking changes are described in NEWS.
create.ian.graph() builds a Gabriel graph and adapts its neighborhoods, preserving
connectivity by default. It returns the initial and final graphs, local scales,
affinities and diagnostics. IAN uses an optional backend; setup currently supports
macOS arm64 and needs Python 3, native Cargo/Rust and Apple's C++ tools.
backend <- build.ian.backend("ian-build") # New directory; explicit one-time setup.
X <- rbind(c(0, 0), c(1, 0), c(2, 0))
fit <- create.ian.graph(X, backend = backend)
if (!fit$complete) stop(fit$error$message)
graph.edges(fit$initial_graph)
graph.edges(fit$final_graph)
fit$diagnostics$connectivity$protected.edges
Keep the returned backend path for later calls. Ordinary package installation does
not compile the optional Rust backend. See vignette("ian-graphs") and
help("build.ian.backend") for setup, reference options, duplicate handling and
resource limits. Successful graph construction does not establish biological or
predictive usefulness.
library(dgraphs)
set.seed(1)
x <- matrix(rnorm(80), ncol = 2)
graph <- create.mknn.graph(x, k = 4)
nrow(graph.edges(graph))
For standard graph examples, choose a type with create.graph():
chain <- create.graph("chain", n = 10, span = 2, labels = 101:110)
circle <- create.graph("circle", n = 20, sampling = "uniform",
edge.length = "chord")
graph.edges(chain)
circle$metadata$coordinates
Other types are "empty", "complete", "cycle", "complete_bipartite",
"star" and "random". Type-specific arguments are documented together in
help("create.graph") and the function guide.
Distances, stored routes, and plotting share the graph representation:
D <- graph.geodesic.distances(chain, vertices = c(1, 10))
paths <- create.path.graph(chain, h.values = c(2, 5))
get.shortest.path(paths[["h_5"]], from = 10, to = 1)
xy <- plot(chain, vertex.values = seq_len(graph.order(chain)))
plot(chain, coordinates = xy, vertex.colors = "navy")
For an end-to-end introduction to graph construction, connectivity repair, parameter sequences, conversion, and diagnostics, run:
vignette("data-derived-graph-workflow", package = "dgraphs")
After installing the development package with build_vignettes = TRUE, open the rendered
vignettes with:
vignette("function-guide", package = "dgraphs")
vignette("synthetic-geometry", package = "dgraphs")
Version 0.3.0 supplies reusable surfaces, curves and point samplers. For example, construct a graph on a quadratic saddle:
surface <- synthetic.quadform(2, 3, list(diag(c(1, -1))))
points <- sample.synthetic.geometry(surface,
synthetic.sampling.uniform.disk(1), n = 100, seed = 4101)
graph <- create.sknn.graph(points$predictors, k = 6, connect.components = FALSE)
The figure uses the exact seed and sample above. Edges carry Euclidean chord lengths, not surface geodesic lengths. Color represents height, from blue (low) through cream to red (high). Its generation recipe also defines the labeled perspective view. The full geometry gallery compares forms, sampling densities, coordinate frames and dimensions.
Geometry-only samples contain coordinates, geometric metadata and reproducible
random-state information. Statistical truth, responses, named recipe registries
and dataset identities remain in geosmooth. The forthcoming geosmooth release imports these geometry helpers from dgraphs. The development Geometry Lab app
is documented in dev/apps/embedding-explorer/README.md; it is not installed
with the R package.
Open saved graph reconstruction benchmarks with dgraphs::explore.graphs().
The optional app dependencies are shiny, bslib, plotly, and digest.
Use explore.graphs(project = run_directory) to open an existing benchmark,
or remember it with register.graph.project(run_directory, name = "My study").
The no-argument launcher presents your saved projects. Scientific data stay
outside the package; opening them does not rerun analysis. grip is needed only
when you explicitly generate a missing weighted layout.
Start with the bundled twelve-point circle comparison:
install.packages(c("shiny", "bslib", "plotly", "digest"))
demo <- system.file("extdata", "graph-explorer-demo", package = "dgraphs")
benchmark <- read.graph.benchmark(demo)
explore.graphs(demo)
Compare k = 2 and k = 4: extra chords increase error relative to exact circle
arcs. Both graph layouts and metric tables are saved, so this demonstration
needs no fitting or project registration. The default display preserves
coordinate proportions, with labeled x/y/z axes. See ?explore.graphs for
project paths and cache controls.