ECV.Rank {randnet}

estimates optimal low rank model for a network

Description

estimates the optimal low rank model for a network

Usage

ECV.Rank(A, max.K, B = 3, holdout.p = 0.1, weighted = TRUE,mode="directed")

Arguments

A

adjacency matrix

max.K

maximum possible rank to check

B

number of replications in ECV

holdout.p

test set proportion

weighted

whether the network is weighted. If TRUE, only sum of squared errors are computed. If FALSE, then treat the network as binary and AUC will be computed along with SSE.

mode

Selectign the mode of "directed" or "undirected" for cross-validation.

Details

AUC is believed to be more accurate in many simulations for binary networks. But the computation of AUC is much slower than SSE, even slower than matrix completion steps.

Note that we do not have to assume the true model is low rank. This function simply finds a best low-rank approximation to the true model.

Value

A list of

sse.rank

rank selection by SSE loss

auc.rank

rank selection by AUC loss

auc

auc sequence for each rank candidate

sse

sse sequence for each rank candidate

Author(s)

Tianxi Li, Elizaveta Levina, Ji Zhu
Maintainer: Tianxi Li [e-mail]

References

T. Li, E. Levina, and J. Zhu. Network cross-validation by edge sampling. Biometrika, 107(2), pp.257-276, 2020.

See Also

ECV.block

Examples

dt <- BlockModel.Gen(30,300,K=3,beta=0.2,rho=0.9,simple=FALSE,power=TRUE)


A <- dt$A


ecv.rank <- ECV.Rank(A,6,weighted=FALSE,mode="undirected")

ecv.rank