Local Geometry in Imbalanced Classification: Simulated Datasets

This dataset contains 180 simulated binary classification datasets designed to study the interaction between class imbalance and local neighborhood geometry. The simulation design combines three imbalance levels, defined as the proportion of the positive class (pi = 0.005, 0.05, 0.5), three local geometric configurations (clean_local, fragmented_minor, majority_intrusion), and 20 replications for each regime-by-imbalance combination.

Each dataset contains 2500 observations, 7 numerical predictors, and a binary class label with levels "pos" and "neg". The positive class is treated as the minority class when pi < 0.5. The simulated configurations are intended to represent different local difficulty mechanisms: well-separated local structure, fragmented minority structure, and majority intrusion into the minority region.