Choosing a Method

At a glance

The methods address related but distinct tasks. Start from the scientific question and the assumed network model rather than treating their scores as interchangeable.

Method	Primary role	Natural use
NETCROP	Model selection by overlapping-network subsampling	Select block-model community count, RDPG or LSM dimension, or a spectral regularizer
ECV	Model selection by holding out edges	Select block-model community count or RDPG dimension
NCV	Community-count selection by holding out nodes	Select an SBM or DCBM community count
DKEST	Spectral-regularizer selection	Tune regularization for spectral clustering
SONNET	Scalable clustering by combining overlapping subnetworks	Fit a large network when a full-network clustering fit is inconvenient

NETCROP, ECV, and NCV are alternative validation strategies. Their loss values come from different holdout constructions and should not be compared as though they were measurements on a common scale. DKEST has the narrower purpose of regularizer tuning. SONNET is principally an estimator, not a substitute for a model-selection criterion; pair it with a justified community count or a selection workflow when that count is unknown.

A common block-model setup

Use the same deterministic network when comparing workflows conceptually. Generator metadata remains available without storing a dense probability matrix.

A <- generate_sbm(
  n = 200,
  K = 3,
  alpha = 0.5,
  beta = 0.08,
  seed = 2026,
  ncores = 1
)
truth <- get_generator_parameters(A)
table(truth$g_true)

## 
##  1  2  3 
## 54 76 70

The generator returns sparse output by default where supported. netOP re-exports Matrix-aware mean(), sum(), diag(), rowMeans(), rowSums(), colMeans(), and colSums(), so ordinary summaries dispatch correctly after library(netOP). Request dense output explicitly only when another tool needs it.

Candidate sets and entry points

For a block model with an unknown community count, a natural introductory candidate set is 1:5:

netcrop_fit <- netcrop_blockmodel(
  A, K_candidates = 1:5,
  nrep = 1, ncores = 1, seed = 1, verbose = FALSE
)

ecv_fit <- ecv_stability_blockmodel(
  A, max_K = 5,
  nrep = 1, ncores = 1, seed = 1, verbose = FALSE
)

ncv_fit <- ncv_stability_blockmodel(
  A, max_K = 5,
  nrep = 1L, ncores = 1, seed = 1, verbose = FALSE
)

ECV and NCV use maximum-size interfaces and evaluate the sequence through that maximum. NETCROP accepts an explicit candidate vector.

For an RDPG or LSM example with true dimension d = 3, use candidate dimensions 1:5:

rdpg_fit <- netcrop_rdpg(
  A, d_candidates = 1:5,
  nrep = 1, ncores = 1, seed = 2, verbose = FALSE
)

ecv_rdpg_fit <- ecv_stability_rdpg(
  A, max_d = 5,
  nrep = 1, ncores = 1, seed = 2, verbose = FALSE
)

lsm_fit <- netcrop_lsm(
  A, d_candidates = 1:5,
  nrep = 1, ncores = 1, seed = 2, verbose = FALSE
)

For regularized spectral clustering, compare NETCROP’s regularizer selector with DKEST, using the same scientifically appropriate candidate grid. For a known K = 3, SONNET provides scalable fitting through sonnet(); its shared- and independent-overlap variants are described on the SONNET help pages.

Reproducibility

Pass seed to randomized generators and fitting or selection routines. Examples use ncores = 1 to behave consistently across operating systems and to keep article builds lightweight. Production runs may use more workers where the function supports them; the individual help page documents how its seed is applied to parallel tasks. Record the package version, candidate set, loss, repetition count, seed, and worker count with reported results.

Glossary

SBM — stochastic block model.
DCBM — degree-corrected stochastic block model.
RDPG — random dot product graph.
LSM — latent-space model.
ASE — adjacency spectral embedding.
ECV — edge cross-validation.
NCV — node cross-validation.
NETCROP — NETwork CRoss-Validation using Overlapping Partitions.

References and disclosures

Use citation("netOP") for machine-readable citations. The method help pages provide method-specific references and implementation disclosures. In particular, netOP’s self-contained ECV implementation is derived from CRAN randnet 1.0, but randnet is not a netOP dependency. File-level licensing and provenance are recorded in inst/COPYRIGHTS.