Site icon R-bloggers

Boosted Configuration (_neural_) Networks for classification

[This article was first published on T. Moudiki's Webpage - R, and kindly contributed to R-bloggers]. (You can report issue about the content on this page here)
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

A few years ago in 2018, I discussed Boosted Configuration (neural) Networks (BCN for multivariate time series forecasting) in this document. Unlike Stochastic Configuration Networks from which they are inspired, BCNs aren’t randomized. Rather, they are closer to Gradient Boosting Machines and Matching Pursuit algorithms; with base learners being single-layered feedforward neural networks – that are actually optimized at each iteration of the algorithm.

The mathematician that you are has certainly been asking himself questions about the convexity of the problem at line 4, algorithm 1 (in the document). As of July 2022, there are unfortunately no answers to that question. BCNs works well empirically, as we’ll see, and finding the maximum at line 4 of the algorithm is achieved, by default, with R’s stats::nlminb. Other derivative-free optimizers are available in R package bcn.

As it will be shown in this document, BCNs can be used for classification. For this purpose, and as implemented in R package bcn, the response (variable to be explained) containing the classes is one-hot encoded as a matrix of probabilities equal to 0 or 1. Then, the classification technique dealing with a one-hot encoded response matrix is similar to the one presented in this post.

6 toy datasets are used for this basic demo of R package bcn: Iris, Wine, Ionosphere, Wisconsin Breast, Digits, Penguins. For each dataset, hyperparameter tuning has already been done. Repeated 5-fold cross-validation was carried out on 80% of the data, for each dataset, and the accuracy reported in the table below is calculated on the remaining 20% of the data. BCN results are compared to Random Forest’s (with default parameters), in order to verify that BCN results are not absurd – it’s not a competition between Random Forest and BCN here.

In the examples, you’ll also notice that BCN’s adjustment to a dataset can be relatively slow when the number of explanatory variables is high (>30, see the digits dataset example, initially with 64 covariates). This is because of line 4, algorithm 1, too: an optimization problem in a high dimensional space is repeated at each iteration of the algorithm.

The future for R package bcn (in no particular order)?

Dataset BCN test set Accuracy Random Forest test set accuracy
iris 100% 93.33%
Wine 97.22% 94.44%
Ionosphere 90.14% 95.77%
Breast cancer 99.12% 94.73%
Digits 97.5% 98.61%
Penguins 100% 100%

Content

0 – Installing and loading packages

Installing bcn From Github:

devtools::install_github("Techtonique/bcn")

# Browse the bcn manual pages
help(package = 'bcn')

Installing bcn from R universe:

# Enable repository from techtonique
options(repos = c(
  techtonique = 'https://techtonique.r-universe.dev',
  CRAN = 'https://cloud.r-project.org'))
  
# Download and install bcn in R
install.packages('bcn')

# Browse the bcn manual pages
help(package = 'bcn')

Loading packages:

library(bcn) # Boosted Configuration networks (only for classification, for now)
library(mlbench) # Machine Learning Benchmark Problems
library(caret)
library(randomForest)
library(pROC)

1 – iris dataset

data("iris")

head(iris)

dim(iris)

set.seed(1234)
train_idx <- sample(nrow(iris), 0.8 * nrow(iris))
X_train <- as.matrix(iris[train_idx, -ncol(iris)])
X_test <- as.matrix(iris[-train_idx, -ncol(iris)])
y_train <- iris$Species[train_idx]
y_test <- iris$Species[-train_idx]

ptm <- proc.time()
fit_obj <- bcn::bcn(x = X_train, y = y_train, B = 10L, nu = 0.335855,
                    lam = 10**0.7837525, r = 1 - 10**(-5.470031), tol = 10**-7,
                    activation = "tanh", type_optim = "nlminb", show_progress = FALSE)
cat("Elapsed: ", (proc.time() - ptm)[3])

plot(fit_obj$errors_norm, type='l')

preds <- predict(fit_obj, newx = X_test)

mean(preds == y_test)

table(y_test, preds)

rf <- randomForest::randomForest(x = X_train, y = y_train)
mean(predict(rf, newdata=as.matrix(X_test)) == y_test)

print(head(predict(fit_obj, newx = X_test, type='probs')))
print(head(predict(rf, newdata=as.matrix(X_test), type='prob')))

2- wine dataset

data(wine)

head(wine)

dim(wine)

set.seed(1234)
train_idx <- sample(nrow(wine), 0.8 * nrow(wine))
X_train <- as.matrix(wine[train_idx, -ncol(wine)])
X_test <- as.matrix(wine[-train_idx, -ncol(wine)])
y_train <- as.factor(wine$target[train_idx])
y_test <- as.factor(wine$target[-train_idx])

ptm <- proc.time()
fit_obj <- bcn::bcn(x = X_train, y = y_train, B = 6L, nu = 0.8715725,
                    lam = 10**0.2143678, r = 1 - 10**(-6.1072786),
                    tol = 10**-4.9605713, show_progress = FALSE)
cat("Elapsed: ", (proc.time() - ptm)[3])

plot(fit_obj$errors_norm, type='l')

preds <- predict(fit_obj, newx = X_test)

mean(preds == y_test)

table(y_test, preds)

rf <- randomForest::randomForest(x = X_train, y = y_train)
mean(predict(rf, newdata=as.matrix(X_test)) == y_test)

print(head(predict(fit_obj, newx = X_test, type='probs')))
print(head(predict(rf, newdata=as.matrix(X_test), type='prob')))

3 – Ionosphere dataset

data("Ionosphere")

head(Ionosphere)

dim(Ionosphere)

Ionosphere$V1 <- as.numeric(Ionosphere$V1)
Ionosphere$V2 <- NULL
set.seed(1234)
train_idx <- sample(nrow(Ionosphere), 0.8 * nrow(Ionosphere))
X_train <- as.matrix(Ionosphere[train_idx, -ncol(Ionosphere)])
X_test <- as.matrix(Ionosphere[-train_idx, -ncol(Ionosphere)])
y_train <- as.factor(Ionosphere$Class[train_idx])
y_test <- as.factor(Ionosphere$Class[-train_idx])

ptm <- proc.time()
fit_obj <- bcn::bcn(x = X_train,
                     y = y_train, B = 50L,
                     nu = 0.5182606,
                     lam = 10**1.323274,
                     r = 1 - 10**(-6.694688),
                     col_sample = 0.7956659,
                     tol = 10**-7,
                     verbose=FALSE,
                     show_progress = FALSE)
cat("Elapsed: ", (proc.time() - ptm)[3])

plot(fit_obj$errors_norm, type='l')

preds <- predict(fit_obj, newx = X_test)

mean(preds == y_test)

table(y_test, preds)

rf <- randomForest::randomForest(x = X_train, y = y_train)
mean(predict(rf, newdata=as.matrix(X_test)) == y_test)

print(head(predict(fit_obj, newx = X_test, type='probs')))
print(head(predict(rf, newdata=as.matrix(X_test), type='prob')))

roc_obj <- pROC::roc(as.numeric(y_test), as.numeric(preds))
pROC::auc(roc_obj)

roc_obj_rf <- pROC::roc(as.numeric(y_test), as.numeric(predict(rf, newdata=as.matrix(X_test))))
pROC::auc(roc_obj_rf)

4 – breast cancer dataset

data("breast_cancer")

head(breast_cancer)

dim(breast_cancer)

set.seed(1234)
train_idx <- sample(nrow(breast_cancer), 0.8 * nrow(breast_cancer))
X_train <- as.matrix(breast_cancer[train_idx, -ncol(breast_cancer)])
X_test <- as.matrix(breast_cancer[-train_idx, -ncol(breast_cancer)])
y_train <- as.factor(breast_cancer$target[train_idx])
y_test <- as.factor(breast_cancer$target[-train_idx])

ptm <- proc.time()
fit_obj <- bcn::bcn(x = X_train, y = y_train, B = 31L, nu = 0.4412851,
                    lam = 10**-0.2439358, r = 1 - 10**(-7), col_sample = 0.5, tol = 10**-2, show_progress = FALSE)
cat("Elapsed: ", (proc.time() - ptm)[3])

plot(fit_obj$errors_norm, type='l')

preds <- predict(fit_obj, newx = X_test)

mean(preds == y_test)

table(y_test, preds)

rf <- randomForest::randomForest(x = X_train, y = y_train)
mean(predict(rf, newdata=as.matrix(X_test)) == y_test)

print(head(predict(fit_obj, newx = X_test, type='probs')))
print(head(predict(rf, newdata=as.matrix(X_test), type='prob')))

roc_obj <- pROC::roc(as.numeric(y_test), as.numeric(preds))
pROC::auc(roc_obj)

roc_obj_rf <- pROC::roc(as.numeric(y_test), as.numeric(predict(rf, newdata=as.matrix(X_test))))
pROC::auc(roc_obj_rf)

5 – digits dataset

data("digits")

head(digits)

dim(digits)

set.seed(1234)
train_idx <- sample(nrow(digits), 0.8 * nrow(digits))
X_train <- as.matrix(digits[train_idx, -ncol(digits)])
X_test <- as.matrix(digits[-train_idx, -ncol(digits)])
y_train <- as.factor(digits$target[train_idx])
X_train <- X_train[, -caret::nearZeroVar(X_train)]
y_test <- as.factor(digits$target[-train_idx])
X_test <- X_test[, colnames(X_train)]

ptm <- proc.time()
fit_obj <- bcn::bcn(x = X_train,
                    y = y_train, B = 50L,
                    nu = 0.6549268,
                    lam = 10**0.4635435,
                    r = 1 - 10**(-7),
                    col_sample = 0.8928518,
                    tol = 10**-5.483609,
                    verbose=FALSE,
                    show_progress = FALSE)
cat("Elapsed: ", (proc.time() - ptm)[3])

plot(fit_obj$errors_norm, type='l')

preds <- predict(fit_obj, newx = X_test)

mean(preds == y_test)

table(y_test, preds)

rf <- randomForest::randomForest(x = X_train, y = y_train)
mean(predict(rf, newdata=as.matrix(X_test)) == y_test)

print(head(predict(fit_obj, newx = X_test, type='probs')))
print(head(predict(rf, newdata=as.matrix(X_test), type='prob')))

6 – Penguins dataset

data("penguins")

penguins_ <- as.data.frame(penguins)

replacement <- median(palmerpenguins::penguins$bill_length_mm, na.rm = TRUE)
penguins_$bill_length_mm[is.na(palmerpenguins::penguins$bill_length_mm)] <- replacement

replacement <- median(palmerpenguins::penguins$bill_depth_mm, na.rm = TRUE)
penguins_$bill_depth_mm[is.na(palmerpenguins::penguins$bill_depth_mm)] <- replacement

replacement <- median(palmerpenguins::penguins$flipper_length_mm, na.rm = TRUE)
penguins_$flipper_length_mm[is.na(palmerpenguins::penguins$flipper_length_mm)] <- replacement

replacement <- median(palmerpenguins::penguins$body_mass_g, na.rm = TRUE)
penguins_$body_mass_g[is.na(palmerpenguins::penguins$body_mass_g)] <- replacement

# replacing NA's by the most frequent occurence
penguins_$sex[is.na(palmerpenguins::penguins$sex)] <- "male" # most frequent

print(summary(penguins_))
print(sum(is.na(penguins_)))

# one-hot encoding for covariates
penguins_mat <- model.matrix(species ~., data=penguins_)[,-1]
penguins_mat <- cbind(penguins_$species, penguins_mat)
penguins_mat <- as.data.frame(penguins_mat)
colnames(penguins_mat)[1] <- "species"

print(head(penguins_mat))
print(tail(penguins_mat))

y <- as.integer(penguins_mat$species)
X <- as.matrix(penguins_mat[,2:ncol(penguins_mat)])

n <- nrow(X)
p <- ncol(X)

set.seed(1234)
index_train <- sample(1:n, size=floor(0.8*n))
X_train <- X[index_train, ]
y_train <- factor(y[index_train])
X_test <- X[-index_train, ]
y_test <- factor(y[-index_train])

ptm <- proc.time()
fit_obj <- bcn::bcn(x = X_train, y = y_train, B = 23, nu = 0.470043,
                    lam = 10**-0.05766029, r = 1 - 10**(-7.905866), tol = 10**-7, 
                    show_progress = FALSE)
cat("Elapsed: ", (proc.time() - ptm)[3])

plot(fit_obj$errors_norm, type='l')

preds <- predict(fit_obj, newx = X_test)

mean(preds == y_test)

table(y_test, preds)

rf <- randomForest::randomForest(x = X_train, y = y_train)
mean(predict(rf, newdata=as.matrix(X_test)) == y_test)

print(head(predict(fit_obj, newx = X_test, type='probs')))
print(head(predict(rf, newdata=as.matrix(X_test), type='prob')))
To leave a comment for the author, please follow the link and comment on their blog: T. Moudiki's Webpage - R.

R-bloggers.com offers daily e-mail updates about R news and tutorials about learning R and many other topics. Click here if you're looking to post or find an R/data-science job.
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.