Add efficiency simulation

This commit is contained in:
Daniel Kapla 2025-10-31 14:58:32 +01:00
parent 491688378c
commit 3724166e8d
6 changed files with 233 additions and 55 deletions

View File

@ -2392,7 +2392,7 @@
@misc{lichess-database,
author = {Duplessis, Thibault},
author = {Thibault Duplessis},
year = {2013},
title = {lichess.org open database},
url = {https://database.lichess.org},

View File

@ -52,10 +52,10 @@ c(X, y) %<-% readRDS("eeg_data_2d.rds")
#'
#' @param X 3D EEG data (preprocessed or not)
#' @param F binary responce `y` as a 3D tensor, every obs. is a 1 x 1 matrix
loo.predict.gmlm <- function(X, y) {
loo.predict.gmlm <- function(X, y, ...) {
unlist(parallel::mclapply(seq_along(y), function(i) {
# Fit with i'th observation removed
fit <- gmlm_tensor_normal(X[ , , -i], as.integer(y[-i]), sample.axis = 3L)
fit <- gmlm_tensor_normal(X[ , , -i], as.integer(y[-i]), sample.axis = 3L, ...)
# Reduce the entire data set
r <- as.vector(mlm(X, fit$betas, modes = 1:2, transpose = TRUE))
@ -75,26 +75,35 @@ loo.predict.gmlm <- function(X, y) {
}, mc.cores = getOption("mc.cores", max(1L, parallel::detectCores() - 1L))))
}
proj.fft <- function(beta1, nr.freq = 5L) {
F <- fft(beta1)
Re(fft(`[<-`(F, head(order(abs(F)), -nr.freq), 0+0i), inverse = TRUE)) / length(F)
}
# perform preprocessed (reduced) and raw (not reduced) leave-one-out prediction
y.hat.3.4 <- loo.predict.gmlm(preprocess(X, 3, 4), y)
y.hat.15.15 <- loo.predict.gmlm(preprocess(X, 15, 15), y)
y.hat.20.30 <- loo.predict.gmlm(preprocess(X, 20, 30), y)
y.hat <- loo.predict.gmlm(X, y)
y.hat.fft <- loo.predict.gmlm(X, y, proj.betas = list(proj.fft, NULL))
# classification performance measures table by leave-one-out cross-validation
(loo.cv <- apply(cbind(y.hat.3.4, y.hat.15.15, y.hat.20.30, y.hat), 2, function(y.pred) {
sapply(c("acc", "err", "fpr", "tpr", "fnr", "tnr", "auc", "auc.sd"),
function(FUN) { match.fun(FUN)(as.integer(y) - 1L, y.pred) })
}))
#> y.hat.3.4 y.hat.15.15 y.hat.20.30 y.hat
#> acc 0.79508197 0.78688525 0.78688525 0.78688525
#> err 0.20491803 0.21311475 0.21311475 0.21311475
#> fpr 0.35555556 0.40000000 0.40000000 0.40000000
#> tpr 0.88311688 0.89610390 0.89610390 0.89610390
#> fnr 0.11688312 0.10389610 0.10389610 0.10389610
#> tnr 0.64444444 0.60000000 0.60000000 0.60000000
#> auc 0.85108225 0.83838384 0.83924964 0.83896104
#> auc.sd 0.03584791 0.03760531 0.03751307 0.03754553
(loo.cv <- apply(
cbind(y.hat.3.4, y.hat.15.15, y.hat.20.30, y.hat, y.hat.fft), 2,
function(y.pred) {
sapply(c("acc", "err", "fpr", "tpr", "fnr", "tnr", "auc", "auc.sd"),
function(FUN) { match.fun(FUN)(as.integer(y) - 1L, y.pred) })
}
))
#> y.hat.3.4 y.hat.15.15 y.hat.20.30 y.hat y.hat.fft
#> acc 0.79508197 0.78688525 0.78688525 0.78688525 0.81147541
#> err 0.20491803 0.21311475 0.21311475 0.21311475 0.18852459
#> fpr 0.35555556 0.40000000 0.40000000 0.40000000 0.33333333
#> tpr 0.88311688 0.89610390 0.89610390 0.89610390 0.89610390
#> fnr 0.11688312 0.10389610 0.10389610 0.10389610 0.10389610
#> tnr 0.64444444 0.60000000 0.60000000 0.60000000 0.66666667
#> auc 0.85194805 0.83838384 0.83924964 0.83896104 0.84646465
#> auc.sd 0.03574475 0.03760531 0.03751754 0.03754553 0.03751864
################################## Tensor SIR ##################################

View File

@ -1,6 +1,8 @@
library(tensorPredictors)
library(parallel)
library(pROC)
suppressPackageStartupMessages({
library(parallel)
library(pROC)
})
#' Mode-Wise PCA preprocessing (generalized (2D)^2 PCA)
#'
@ -43,28 +45,6 @@ auc.sd <- function(y.true, y.pred) {
}
# # unified API for all reduction procedures
# GMLM <- list(
# fit = function(X, y) tensorPredictors::gmlm_tensor_normal(X, as.integer(y), sample.axis = 4L),
# reduce = function(X, fit) mlm(X, fit$betas, 1:3, TRUE),
# applicable = function(X) TRUE
# )
# TSIR <- list(
# fit = function(X, y) tensorPredictors::TSIR(X, y, c(1L, 1L, 1L), sample.axis = 4L),
# reduce = function(X, fit) mlm(X, fit, 1:3, TRUE),
# applicable = function(X) TRUE
# )
# KPIR_LS <- list(
# fit = function(X, y) {
# if (any(dim(X)[-4] > dim(X)[4])) {
# stop("Dimensions too big")
# }
# tensorPredictors::kpir.ls(X, as.integer(y), sample.axis = 4L)
# },
# reduce = function(X, fit) if (is.null(fit)) NA else mlm(X, fit$alphas, 1:3, TRUE),
# applicable = function(X) all(dim(X)[1:3] <= dim(X)[4])
# )
#' Leave-one-out prediction using TSIR
#'
#' @param method reduction method to be applied
@ -98,6 +78,12 @@ loo.predict <- function(method, X, y, ...) {
}, mc.cores = getOption("mc.cores", max(1L, parallel::detectCores() - 1L))))
}
# "Projects" a sequence to its first `nr.freq` frequency components
proj.fft <- function(sequence, nr.freq = 5L) {
F <- fft(sequence)
Re(fft(`[<-`(F, head(order(abs(F)), -nr.freq), 0+0i), inverse = TRUE)) / length(F)
}
# Load full EEG dataset (3D tensor for each subject)
c(X, y) %<-% readRDS("eeg_data_3d.rds")
@ -110,21 +96,22 @@ y.hat.3.4 <- loo.predict(gmlm_tensor_normal, preprocess(X, 3, 4, 3), y)
y.hat.15.15 <- loo.predict(gmlm_tensor_normal, preprocess(X, 15, 15, 3), y)
y.hat.20.30 <- loo.predict(gmlm_tensor_normal, preprocess(X, 20, 30, 3), y)
y.hat <- loo.predict(gmlm_tensor_normal, X, y)
y.hat.fft <- loo.predict(gmlm_tensor_normal, X, y, proj.betas = list(proj.fft, NULL, NULL))
# classification performance measures table by leave-one-out cross-validation
(loo.cv <- apply(cbind(y.hat.3.4, y.hat.15.15, y.hat.20.30, y.hat), 2, function(y.pred) {
(loo.cv <- apply(cbind(y.hat.3.4, y.hat.15.15, y.hat.20.30, y.hat, y.hat.fft), 2, function(y.pred) {
sapply(c("acc", "err", "fpr", "tpr", "fnr", "tnr", "auc", "auc.sd"),
function(FUN) { match.fun(FUN)(as.integer(y) - 1L, y.pred) })
}))
#> y.hat.3.4 y.hat.15.15 y.hat.20.30 y.hat
#> acc 0.83606557 0.80327869 0.80327869 0.79508197
#> err 0.16393443 0.19672131 0.19672131 0.20491803
#> fpr 0.31111111 0.33333333 0.33333333 0.35555556
#> tpr 0.92207792 0.88311688 0.88311688 0.88311688
#> fnr 0.07792208 0.11688312 0.11688312 0.11688312
#> tnr 0.68888889 0.66666667 0.66666667 0.64444444
#> auc 0.88051948 0.86984127 0.86926407 0.86810967
#> auc.sd 0.03118211 0.03254642 0.03259186 0.03295883
#> y.hat.3.4 y.hat.15.15 y.hat.20.30 y.hat y.hat.fft
#> acc 0.83606557 0.80327869 0.80327869 0.79508197 0.79508197
#> err 0.16393443 0.19672131 0.19672131 0.20491803 0.20491803
#> fpr 0.31111111 0.33333333 0.33333333 0.35555556 0.33333333
#> tpr 0.92207792 0.88311688 0.88311688 0.88311688 0.87012987
#> fnr 0.07792208 0.11688312 0.11688312 0.11688312 0.12987013
#> tnr 0.68888889 0.66666667 0.66666667 0.64444444 0.66666667
#> auc 0.88051948 0.86984127 0.86926407 0.86810967 0.86810967
#> auc.sd 0.03118211 0.03254642 0.03259186 0.03295883 0.03354029
################################## Tensor SIR ##################################

182
sim/sim_efficiency.R Normal file
View File

@ -0,0 +1,182 @@
library(tensorPredictors)
# Set PRNG seed to the first 4 digits of the golden ratio for reproducability
set.seed(1618L, "Mersenne-Twister", "Inversion", "Rejection")
### Simulation configuration
reps <- 100 # number of simulation replications
sample.sizes <- c(100, 200, 300, 500, 750) # sample sizes `n`
# Parameterize the "true" reductions on the dimension
gen.beta <- function(pj) {
as.matrix((-1)^seq_len(pj))
}
# the precision matrices are simply diag(pj)
# sampling from the conditional matrix normal `X | Y = y ~ N(mu(y), I_{p1 p2})`
sample.data <- function(sample.size, betas, Omegas, eta1 = 0) {
# responce is a standard normal variable
y <- rnorm(sample.size)
# F(y) is identical to y, except its a tensor (last axis is sample axis)
F <- array(y, dim = c(mapply(ncol, betas), sample.size))
# sample predictors from tensor normal X | Y = y (last axis is sample axis)
sample.axis <- length(betas) + 1L
Deltas <- Map(solve, Omegas) # normal covariances
mu_y <- mlm(mlm(F, betas) + as.vector(eta1), Deltas) # conditional mean
X <- mu_y + rtensornorm(sample.size, 0, Deltas, sample.axis) # responses X
list(X = X, F = F, y = y, sample.axis = sample.axis)
}
# Open simulation CSV log file
log.name <- format(Sys.time(), "sim_efficiency-%Y%m%dT%H%M.csv")
log.file <- file(log.name, "w")
# Counts new number of writes purely here to write the CSV header the first time
log.writes <- 0L
# Setting p1 = p2 = pj (note, in the paper `p = p1 p2`)
mode.dims <- round(1.2^unique(round(logb(2:200, 1.2))))
for (pj in mode.dims) {
betas.true <- list(gen.beta(pj), gen.beta(pj))
B.true <- kronecker(betas.true[[2]], betas.true[[1]])
Omegas.true <- list(diag(pj), diag(pj))
for (sample.size in sample.sizes) {
sim <- sapply(seq_len(reps), function(.) {
c(X, F, y, sample.axis) %<-% sample.data(sample.size, betas.true, Omegas.true)
ds.lm <- tryCatch({
unname(lm.fit(t(`dim<-`(X, c(pj^2, sample.size))), drop(F))$coefficients)
dist.subspace(B.true, B.lm, normalize = TRUE)
}, error = function(.) NA)
c(., betas.vec, Omegas.vec) %<-% gmlm_tensor_normal(`dim<-`(X, c(pj^2, sample.size)), drop(F))
c(., betas.gmlm, Omegas.gmlm) %<-% gmlm_tensor_normal(X, F)
c(., betas.mani, Omegas.mani) %<-% gmlm_tensor_normal(X, F,
proj.Omegas = rep(list(function(O) { diag(mean(diag(O)), nrow(O)) }), 2)
)
ds.vec <- dist.subspace(B.true, betas.vec[[1]], normalize = TRUE)
ds.gmlm <- dist.subspace(betas.true, betas.gmlm, normalize = TRUE) # equiv to R> dist.subspace(B.true, B.gmlm)
ds.mani <- dist.subspace(betas.true, betas.mani, normalize = TRUE)
c(lm = ds.lm, vec = ds.vec, gmlm = ds.gmlm, mani = ds.mani)
})
sim <- as.data.frame(t(sim))
sim$sample.size <- sample.size
sim$pj <- pj
# boxplot(t(sim))
# summary(t(sim))
# Append current simulation results to log-file
write.table(sim, file = log.file, sep = ",",
row.names = FALSE, col.names = (log.writes <- log.writes + 1L) < 2L
)
# print progress
cat(sprintf("mode dim (%d): %d/%d - sample size (%d): %d/%d\n",
pj, which(pj == mode.dims), length(mode.dims),
sample.size, which(sample.size == sample.sizes), length(sample.sizes)
))
}
}
close(log.file)
# Read simulation data back in
sim <- read.csv(log.name)
# with(aggregate(sim, . ~ sample.size + pj, mean), {
# plot(range(pj), range(c(vec, gmlm, mani)), type = "n",
# main = "Simulation -- Efficiency Gain",
# xlab = expression(tilde(p)),
# ylab = expression(d(B, hat(B)))
# )
# for (sz in sort(unique(sample.size))) {
# i <- order(pj)[sample.size == sz]
# lines(pj[i], vec[i], type = "b", pch = 16, col = sz %/% 100, lty = 1)
# lines(pj[i], gmlm[i], type = "b", pch = 16, col = sz %/% 100, lty = 2)
# lines(pj[i], mani[i], type = "b", pch = 16, col = sz %/% 100, lty = 3)
# }
# sd <- aggregate(sim, . ~ sample.size + pj, sd)
# })
with(merge(
aggregate(sim[names(sim) != "lm"], . ~ sample.size + pj, mean),
aggregate(sim[names(sim) != "lm"], . ~ sample.size + pj, sd),
by = c("sample.size", "pj"),
suffixes = c("", ".sd"),
all = FALSE
), {
plot(range(pj), range(c(vec, gmlm, mani)), type = "n",
main = "Simulation -- Efficiency Gain",
xlab = expression(tilde(p)),
ylab = expression(d(B, hat(B)))
)
# colors <- c("#cf7d03ff", "#002d8d", "#006e18")
# col.idx <- 0L
lty.idx <- 0L
for (sz in sort(unique(sample.size))) {
i <- order(pj)[(sample.size == sz)[order(pj)]]
# polygon(c(pj[i], rev(pj[i])), c(lm[i] + lm.sd[i], rev(lm[i] - lm.sd[i])),
# col = paste0("#cf7d03", "50"), border = NA
# )
polygon(c(pj[i], rev(pj[i])), c(vec[i] + vec.sd[i], rev(vec[i] - vec.sd[i])),
col = paste0("#b30303", "50"), border = NA
)
polygon(c(pj[i], rev(pj[i])), c(gmlm[i] + gmlm.sd[i], rev(gmlm[i] - gmlm.sd[i])),
col = paste0("#002d8d", "50"), border = NA
)
polygon(c(pj[i], rev(pj[i])), c(mani[i] + mani.sd[i], rev(mani[i] - mani.sd[i])),
col = paste0("#006e18", "50"), border = NA
)
}
lty.idx <- 1L
for (sz in sort(unique(sample.size))) {
i <- order(pj)[(sample.size == sz)[order(pj)]]
# lines(pj[i], lm[i], type = "b", pch = 16, col = "#cf7d03", lty = lty.idx, lwd = 2)
lines(pj[i], vec[i], type = "b", pch = 16, col = "#b30303", lty = lty.idx, lwd = 2)
lines(pj[i], gmlm[i], type = "b", pch = 16, col = "#002d8d", lty = lty.idx, lwd = 2)
lines(pj[i], mani[i], type = "b", pch = 16, col = "#006e18", lty = lty.idx, lwd = 2)
lty.idx <- lty.idx + 1L
}
})
# unname(lm.fit(t(`dim<-`(X, c(pj^2, sample.size))), drop(F))$coefficients)
# unname(lm(drop(F) ~ t(`dim<-`(X, c(pj^2, sample.size))) - 1)$coefficients)
# require(utils)
# set.seed(129)
# n <- 7 ; p <- 2
# X <- matrix(rnorm(n * p), n, p) # no intercept!
# y <- rnorm(n)
# w <- rnorm(n)^2
# str(lmw <- lm.wfit(x = X, y = y, w = w))
# str(lm. <- lm.fit (x = X, y = y))
# if(require("microbenchmark")) {
# mb <- microbenchmark(lm(y~X), lm.fit(X,y), .lm.fit(X,y))
# print(mb)
# boxplot(mb, notch=TRUE)
# }

View File

@ -114,9 +114,9 @@ gmlm_tensor_normal <- function(X, F, sample.axis = length(dim(X)),
Sigmas[[j]] <- Sigmas[[j]] + diag(0.2 * min_max[2], nrow(Sigmas[[j]]))
}
}
# Compute (unconstraint but regularized) Omega_j as covariance inverse
# Compute (unconstraint) Omega_j's as covariance inverse
Omegas[[j]] <- solve(Sigmas[[j]])
# Project Omega_j to the Omega_j's manifold
# Project Omega_j's to their manifolds
if (is.function(proj_j <- proj.Omegas[[j]])) {
Omegas[[j]] <- proj_j(Omegas[[j]])
# Reverse computation of `Sigma_j` as inverse of `Omega_j`

View File

@ -44,14 +44,14 @@ kpir.ls <- function(X, Fy, max.iter = 20L, sample.axis = 1L,
}, modes, dim(Fy)[modes])
### Step 2: iterate per mode (axis) least squares estimates
### Step 2: iterate per mode (axis) least squares estimates
for (iter in seq_len(max.iter)) {
# Invoke logger for previous iterate
if (is.function(logger)) {
logger("ls", iter - 1L, alphas)
}
# cyclic iterate over modes
for (j in seq_along(modes)) {
# least squares solution for `alpha_j | alpha_i, i != j`