ptf (version 0.0.1)

ptf_sparse: Fit a Probit RESCAL model (sparse representation).

Description

Fit a Probit RESCAL model (sparse representation).

Usage

ptf_sparse(df, n, k, r = 0, max_iter = 500, tol = 1e-08,
  tol_M = 1e-05, iter_M_max = 2, print_option = TRUE)

Arguments

df

a four-column dataframe with columns (1) ent1 (2) ent2 (3) relation (4) true, which is an indicator of whether there is such a relation from ent1 to ent2

n

number of entities

k

number of relations

r

decomposition rank

max_iter

max number of iterations

tol

tolerance of absolute change in ,elihood

tol_M

tolerance of absolute change in the M step

iter_M_max

max number of iterations for M step

print_option

whether print log-likelihood for each iteration or not

Value

fitted parameters

References

Ye Liu, 2021. Computational Methods for Complex Models with Latent Structure. PhD thesis with link at https://repository.lib.ncsu.edu/bitstream/handle/1840.20/37507/etd.pdf?sequence=1

Examples

# NOT RUN {
n <- 20
k <- 10
r <- 3
A <- matrix(runif(n*r,0,1),n,r)
mean.k <- runif(k,-0.1,-0.01)
W <- lapply(1:k, function(i) matrix(rnorm(r*r,mean.k[i],1),r,r))
Z <- lapply(1:k, function(i) A%*%W[[i]]%*%t(A))
df <- data.frame(expand.grid(0:(n-1),0:(n-1),0:(k-1)))
df <-  df[sample(nrow(df),2000),]
colnames(df) <- c('ent1','ent2','relation')
awa <- function(row,A,W){
  inx1 <- row[1] + 1
  inx2 <- row[2] + 1
  rel <- row[3] + 1
  out <- A[inx1,] %*% W[[rel]] %*% A[inx2,]
}
z <- apply(df,1,function(row) awa(row,A,W))
df$true <- ifelse(z>0,1,0)
result2 <- ptf_sparse(df,n,k,r)

# }