Synthetic Minority Oversampling for Imbalanced Time Series Classification Based on Path Signature

Imbalanced class distributions hinder time series classifiers by underrepresenting rare yet important events. We introduce Path Signature Synthetic Time-series Oversampling (PSSTO), a structure-preserving oversampling method that operates in path signature space to synthesize informative minority samples while pruning low-quality ones. Across 12 public datasets, PSSTO with a random forest improves classification over conventional resampling approaches on average. Pairwise Wilcoxon signed-rank tests against these approaches indicate statistically significant gains. Compared with time series-specific oversamplers, PSSTO with random forest attains the best averages on F1, G-mean, and AUC compared to the strongest alternative. These results show that structure-preserving oversampling in signature space is an effective and broadly applicable remedy for imbalanced time-series classification.