network.mixing {randnet}

estimates network connection probability by network mixing

Description

estimates network connection probability by network mixing of Li and Le (2021).

Usage

network.mixing(A,index=NULL,rho = 0.1,max.K=15,dcsbm=TRUE, usvt=TRUE,ns=FALSE,
                           lsm=FALSE,lsm.k=4,trace=FALSE)

Arguments

A

adjacency matrix

index

a pre-specified hold-out set. If NULL, the set will be randomly generated according to rho.

rho

hold-out proportion as validation entries. Only effective when index is NULL.

max.K

the maximum number of blocks used for the block model approximation (see details).

dcsbm

whether to include the DCSBM as components, up to max.K. By default, the method will include it.

usvt

whether to include the USVT as a component. By default, the method will include it.

ns

whether to include the neighborhood smoothing as a component.

lsm

whether to include the gradient estimator of the latent space model as a component.

lsm.k

the dimension of the latent space. Only effective if lsm is TRUE.

trace

whether to print the model fitting progress.

Details

The basic version of the mixing estimator will include SBM and DCSBM estimates with the number of blocks from 1 to max.K. Users could also specify whether to include additional USVT, neighborhood smoothing and latent space model estimators. If NNL (non-negative linear), exponential, or ECV is used, the mixing is usually robust for a reasonable range of max.K and whether to include the other models. The linear mixing, however, is vulnerable for a large number of base estimates. The NNL is our recommended method. USVT is also recommended. the neighborhood smoothing and latent space model are slower, so are not suitable for large networks. Details can be found in Li and Le (2021).

Value

a list of

linear.Phat

estimated probability matrix by linear mixing

linear.weight

the weights of the indivdiual models in linear mixing

nnl.Phat

estimated probability matrix by NNL mixing

nnl.weight

the weights of the indivdiual models in NNL mixing

exp.Phat

estimated probability matrix by exponential mixing

exp.weight

the weights of the indivdiual models in exponential mixing

ecv.Phat

estimated probability matrix by ECV mixing (only one nonzero)

ecv.weight

the weights of the indivdiual models in ECV mixing (only one nonzero)

model.names

the names of all individual models, in the same order as the weights

Author(s)

Tianxi Li and Can M. Le

Maintainer: Tianxi Li <[e-mail]>

References

T. Li and C. M. Le, Network Estimation by Mixing: Adaptivity and More. arXiv preprint arXiv:2106.02803, 2021.

Examples

dt <- RDPG.Gen(n=500,K=5,directed=TRUE)

A <- dt$A

fit <- network.mixing(A)
fit$model.names

fit$nnl.weight