Uncertainty Guided Oversampling (UGO, 2026)

UGO (Saglam & Cengiz, 2026) is not a resampling method on its own. It is a classifier-guided procedure that uses a base oversampling method (type) and a classifier: the classifier decides which samples are informative, which class needs more samples and when to stop.

Usage

UGO(
  x,
  y,
  type = c("RWO", "SMOTE", "ROS", "MCSMOTE"),
  classifier = "RF",
  comparison = c("anova", "kruskal"),
  prop_generate = 0.1,
  prop_sampling = 0.5,
  k_max = 5,
  k_noise = 5,
  pen_noise = 1,
  k_after = 1,
  lambda_max = 2.5,
  alpha = 0.05,
  classifier_args = list(),
  generator_args = list(),
  verbose = FALSE,
  ...
)

Arguments

x: feature matrix or dataframe.
y: a factor class variable. Can have more than two classes.
type: base oversampling method used to generate synthetic samples around the selected informative samples. One of "RWO", "SMOTE", "ROS" or "MCSMOTE". Default is "RWO".
classifier: classifier used to calculate the uncertainty. One of the built-in classifiers "CART" (rpart), "RF" (randomForest), "LogReg" (logistic regression, one-vs-rest for more than two classes), "KNN" (FNN), "NaiveBayes" (Gaussian), "LDA" or "NNET" (nnet), or a function function(x_train, y_train, x_test, ...) that returns a matrix of class probabilities for the rows of x_test, with one column per level of y_train. The old names "rpart" and "randomForest" are still accepted. UGO is classifier specific, so the classifier that will be used for the final model should be used here. Default is "RF".
comparison: test used to compare the average uncertainty of the classes: "anova" (Welch's one-way test) or "kruskal" (Kruskal-Wallis). Generation stops when classes are not significantly different.
prop_generate: proportion of the sample size generated at each iteration (\(\lambda_{step}\) in the paper). Default is 0.10.
prop_sampling: proportion of the samples of the selected class that are chosen as the most uncertain samples to generate data from (\(\lambda_{select}\) in the paper). Default is 0.5.
k_max: number of nearest neighbors for type = "SMOTE". Default is 5. If NULL, ceiling(n / number of selected samples).
k_noise: number of neighbors for the initial noise detection (\(k_{g-Pre}\) in the paper). Default is 5. If 0, no sample is treated as noise.
pen_noise: multiplier of the threshold of the noise detection. A sample is noise when fewer than k_noise / K * pen_noise of its neighbors share its class, where \(K\) is the number of classes. Default is 1.
k_after: number of neighbors for the detection of noise among the generated samples (\(k_{g-After}\) in the paper). Default is 1.
lambda_max: the algorithm stops when the data size exceeds lambda_max * K * n (\(\phi\) in the paper). Default is 2.5.
alpha: significance level of the comparison test used as the stopping criterion. Default is 0.05.
classifier_args: list of additional arguments for the classifier (e.g. list(k = 7) for "KNN").
generator_args: list of additional arguments for the base method type. Only thresh_qntl, tol and max_iter of MCSMOTE are used.
verbose: print the progress. Default is FALSE.
...: not used.

Value

a list with resampled dataset.

x_new: Resampled feature matrix.
y_new: Resampled target variable.
x_syn: Generated synthetic feature data.
y_syn: Generated synthetic label data.
i_noise: Logical vector of samples detected as noise.
stop_reason: Why the algorithm stopped: "no_difference" (the uncertainties of the classes are not significantly different), "max_size", "too_much_noise" or "no_uncertainty".
n_iter: Number of completed generation iterations.
unc_mean_initial, unc_mean_final: Average uncertainty of each class (noise excluded) before the first and at the last iteration.
unc_rank_mean_initial, unc_rank_mean_final: Average rank of the uncertainty of each class before the first and at the last iteration.

Details

Year: 2026. UGO (Saglam & Cengiz, 2026) is an iterative oversampling method which uses the uncertainty of a classifier as in active learning. At each iteration, a classifier is trained on the current data and the modified margin uncertainty (the highest probability of the other classes minus the probability of the true class) is calculated for the original samples that are not noise. The class with the highest average uncertainty is selected and prop_generate * n synthetic samples are generated around its prop_sampling most uncertain samples with the base method type. Generated samples that fall in the region of another class are relocated to the nearest selected sample. The algorithm stops when the average uncertainties of the classes are not significantly different, so the number of synthetic samples does not need to be specified. Samples in class overlap regions (noise) are excluded from generation.

References

Saglam, F., & Cengiz, M. A. (2026). Uncertainty guided oversampling: A classifier-specific approach for effective handling of class imbalance in machine learning. Knowledge-Based Systems, 342, 115886.

Author

Fatih Saglam, saglamf89@gmail.com

Examples


set.seed(1)
x <- rbind(matrix(rnorm(600, 3, 1), ncol = 2, nrow = 300),
           matrix(rnorm(60, 5, 1), ncol = 2, nrow = 30))
y <- as.factor(c(rep("negative", 300), rep("positive", 30)))

plot(x, col = y)


# resampling
if (requireNamespace("randomForest", quietly = TRUE)) {
  m <- UGO(x = x, y = y, type = "SMOTE", classifier = "RF")
  m$stop_reason

  plot(m$x_new, col = m$y_new)
}


# built-in classifiers that need no extra package
m <- UGO(x = x, y = y, type = "SMOTE", classifier = "LDA")
m$stop_reason
#> [1] "no_difference"

# any classifier that returns class probabilities can be used
f_logistic <- function(x_train, y_train, x_test, ...) {
  dat <- data.frame(x_train, y = y_train)
  model <- glm(y ~ ., data = dat, family = binomial())
  p1 <- predict(model, newdata = data.frame(x_test), type = "response")
  cbind(1 - p1, p1)
}
m <- UGO(x = x, y = y, type = "ROS", classifier = f_logistic)
table(m$y_new)
#> 
#> negative positive 
#>      300       63