TRACOS: realistic synthetic data generation via chaotic trajectory modeling in high-dimensional imbalanced datasets

Background Imbalanced data distribution is a pervasive challenge in supervised learning, frequently causing models to favor the majority class. Conventional oversampling techniques, such as Synthetic Minority Over-sampling Technique (SMOTE) and its variants, often struggle to preserve the structural complexity of minority class instances, leading to overlapping classes and reduced generalization. Methods To address these limitations, this article introduces TRAjectorybased Chaotic Oversampling Strategy (TRACOS), a novel hybrid oversampling approach. The method first constructs smooth trajectories using adaptive spline interpolation to accurately capture the nonlinear manifold of the minority class. Subsequently, it applies controlled randomness via logistic chaotic maps to generate realistic synthetic samples along these trajectories, enhancing diversity while strictly maintaining structural fidelity. Results TRACOS was rigorously evaluated on 35 multidimensional and highly imbalanced datasets from the Knowledge Extraction based on Evolutionary Learning (KEEL) repository using four distinct classifiers (K-Nearest Neighbors, Decision Tree, Naive Bayes, and Random Forest). The proposed method was benchmarked against seven widely used oversampling techniques, including SMOTE, Borderline-SMOTE variants, SMOTE with Density estimation (SMOTE_D), SMOTE with Rough Set Theory (SMOTE_RSB), and Adaptive Semi-Unsupervised Weighted Oversampling (A_SUWO), alongside the original imbalanced datasets. Based on Area Under the Curve (AUC) and Geometric Mean (G-Mean) metrics under a 5-fold cross-validation setting, results demonstrate that TRACOS consistently achieves superior classification performance. Notably, under the Random Forest model, TRACOS achieved an average G-Mean of 0.8243 and ranked first in 22 out of 35 datasets, exhibiting the highest average scores and the lowest performance variance across all experiments. Discussion The dual mechanism of TRACOS capturing intrinsic data geometry through spline trajectories while safely promoting diversity via chaotic dynamics proves highly effective. It offers a robust, adaptable, and statistically superior solution for real-world classification applications characterized by severe class imbalance.