ssVN: Semi-superved Vertex Nomination

In Noobivsho/ssClust: GMM for semi-supervised case a-la Mclust

Description

Uses information about RED and NOT RED memberships to nominate RED vertices.

Usage

ssVN(adj, embeddingDim, knownRed, knownNotRed, 
  initializationStrategy = "kpp", Grange = c(2:5))

Arguments

adj	Adjacency matrix
embeddingDim	Scalar dimension to embed the adjacency matrix into
knownRed	indices of the vertices that are known to be red
knownNotRed	indices of the vertices that are known NOT to be red
initializationStrategy	currently only accepts "kpp" for semi-supervised k-means++
Grange	number of classes to consider (2 minimum for red and a single not red class)

Value

redRanking	vector of posterior probabilities of being red. Supervised points are easy money (0 or 1, respectively)
redDist	vector of mahalanobis distance to red center
ss	object as returned by the call to ssClust (see documentation in this package)

Author(s)

Jordan Yoder

Examples

library(ssClust)
library(igraph)
n=500
MCR=100
numNumKnownNotRed <- 5
numKnownNotRedPerJ <- 20
j=6
numKnownNotRed <- (j-1)*numKnownNotRedPerJ

###### FORM ADJ MATRIX #####
#matching Li's simulation parameters
rho <- c(0.4,0.3,.3)
B1 <- matrix(c(0.5,0.3,.4,
               0.3,0.8,.6,
               .4,0.6, .3),  nrow = 3,ncol=3)
B2 <- matrix(.5 ,nrow = 3,ncol=3)
if(n==10)
  nu=1
if(n==500)
  nu = .3
if(n==10000)
  nu = .1
B = nu*B1 + (1-nu)*B2
blockSizes <- n*rho

#simulate graph
A.igraph<- sbm.game(n, pref.matrix = B, block.sizes = blockSizes, 
  directed = FALSE, loops = FALSE)
###### END FORM ADJ MATRIX ########

#shovel into obj we need
A <- NULL
A$adj <- get.adjacency(A.igraph)

#tau is actual labels
A$tau <- rep(1,blockSizes[1])
A$tau <- c(A$tau, rep(2, blockSizes[2]))
A$tau <- c(A$tau, rep(3, blockSizes[3]))
redGroup = min(A$tau)


#semi-supervise
if(n ==10)
  knownRed <- sample(which(A$tau==redGroup),4)
if(n ==500)
  knownRed <- sample(which(A$tau==redGroup),20)
if(n ==10000)
  knownRed <- sample(which(A$tau==redGroup),40)

#more semi-supervision
knownNotRed <- sample(which(A$tau!=1),
                      numKnownNotRed)

#cluster and nominate
ss = ssVN(adj=A$adj,
          embeddingDim=3,
          knownRed = knownRed,
          knownNotRed = knownNotRed,
          initializationStrategy = "kpp",
          Grange=2:5)
library(ROCR)
pred = prediction(ss$redRanking, A$tau==redGroup)
perf = performance(pred, 'tpr','fpr')
plot(perf)

Noobivsho/ssClust documentation built on Aug. 10, 2019, 5:47 a.m.